CN109190455A - Black smoke vehicle recognition methods based on Gaussian Mixture and autoregressive moving-average model - Google Patents
Black smoke vehicle recognition methods based on Gaussian Mixture and autoregressive moving-average model Download PDFInfo
- Publication number
- CN109190455A CN109190455A CN201810789111.3A CN201810789111A CN109190455A CN 109190455 A CN109190455 A CN 109190455A CN 201810789111 A CN201810789111 A CN 201810789111A CN 109190455 A CN109190455 A CN 109190455A
- Authority
- CN
- China
- Prior art keywords
- feature
- value
- pixel
- black smoke
- vehicle
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
- G06V10/507—Summing image-intensity values; Histogram projection analysis
-
- 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
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Abstract
Black smoke vehicle recognition methods based on Gaussian Mixture and autoregressive moving-average model includes the following steps: that (1) detects vehicle movement target using gauss hybrid models from traffic surveillance videos；(2) three kinds of features of vehicle key area, including Haar-like feature, co-occurrence matrix gradient orientation histogram feature and local binary patterns Fourier's histogram feature are extracted；(3) it is modeled using continuous multiple frames of the autoregressive moving-average model to every kind of feature, obtains three different models；(4) for new vehicle target, three models are respectively used to the classification for three kinds of features that vehicle key area extracts, in conjunction with the classification results of different characteristic and the comprehensive analysis of continuous multiple frames, whether have black smoke vehicle to judge current video section.The present invention can greatly save the manpower and financial resources of conventional method consumption, be conducive to the acquisition and preservation of evidence, do not influence normal traffic, can effectively improve law enforcement efficiency.
Description
Technical field
The present invention relates to computer visions and mode identification technology, more particularly to based on Gaussian Mixture and autoregression
The black smoke vehicle recognition methods of moving average model.
Background technique
The more serious diesel vehicle of damage ratio is also known as black smoke vehicle, and dense black smoke, pollution one are usually had at gas vent
It is directly the emphasis of Motor Vehicle Emission Control, finds the black smoke vehicle that travels on road in time, and is done by environmental protection administration further
Reason, it will help reduce automobile pollution, improve air quality, reduce its harm to human body.
Many cities are still detected on road by the way of road inspection, reports and manual video monitoring at present
Black smoke vehicle, it is time-consuming and laborious, it also will affect normal traffic circulation, be unfavorable for the preservation of the illegal evidence of associated vehicle.It is now random
How the continuous development of computer vision technique and mode identification technology detects black smoke vehicle automatically by the way of video analysis
Hot spot is increasingly becomed, is highly studied.
The intelligent black smoke vehicle detection method based on video analysis that the present invention provides a kind of, the invention is by gauss hybrid models
It is combined with autoregressive moving-average model, while devising Haar-like feature, co-occurrence matrix gradient orientation histogram is special
(CoHOG) and local binary patterns Fourier histogram feature (LBP-HF) are levied, the fusion of three kinds of features increases the knowledge of black smoke vehicle
Other accuracy rate, reduces rate of false alarm.
Summary of the invention
For problem above, the present invention provides the black smoke vehicle identification side based on Gaussian Mixture and autoregressive moving-average model
Method can integrate different types of static nature, and introduce the behavioral characteristics that autoregressive moving-average model portrays key area,
There is very high robustness to camera shake and vehicle shadow, the interference of noise can be reduced, further increase the identification of black smoke vehicle
Rate reduces rate of false alarm.For this purpose, the present invention provides the black smoke vehicle based on Gaussian Mixture and autoregressive moving-average model
Recognition methods, specific step is as follows, it is characterised in that:
(1) vehicle movement target is detected from traffic surveillance videos using gauss hybrid models；
(2) three kinds of features of vehicle key area, including Haar-like feature, co-occurrence matrix gradient direction histogram are extracted
Figure feature and local binary patterns Fourier's histogram feature；
(3) it is modeled using continuous multiple frames of the autoregressive moving-average model to every kind of feature, obtains three different moulds
Type；
(4) for new vehicle target, three models are respectively used to three kinds of features that vehicle key area extracts
Whether classification, in conjunction with the classification results of different characteristic and the comprehensive analysis of continuous multiple frames, have black smoke vehicle to make current video section
Judgement.
Further, vehicle movement target packet is detected from traffic surveillance videos using gauss hybrid models in step (1)
Include following steps:
(11) gray level image is converted by each frame image, and is divided into the fritter of 3x3 pixel, each block of pixels is established
Gauss model reduces the interference of noise to improve background modeling speed and stability；
(12) model initialization, by taking one block of pixels M of certain frame image as an example, the initial model of the block of pixels utilizes preceding N frame
The block of pixels sequence of sequence image is established, and { x is denoted as1,x2,...,xN, take the average gray μ of block of pixels0And varianceJust
The mean value and variance of the 1st Gaussian Profile of beginningization, i.e.,
Wherein, I (xi,yj, k) and indicate pixel value of the block of pixels in kth frame image at position (x, y)；
(13) illustrate the update method of background model by taking N+1 frame as an example,
If the gray value x of block of pixelsN+1Meet | xN+1-μi,N|≤2.5σi,N, then the block of pixels is divided with corresponding K Gauss
Cloth matching, parameter update is as follows,
Wherein, β and ρ is the learning rate of Gaussian Profile mean value and variance respectively,Indicate Gaussian probability-density function, θ and γ
For fixed value, for the value range of regularized learning algorithm rate, xN+1For the average gray of N+1 frame block of pixels, μi,NAnd σi,N 2For
N+1 frame, the mean value and variance of i-th of Gauss model；
If the gray value x of the block of pixelsN+1It is mismatched with corresponding K Gaussian Profile, then parameter μi,NAnd σi,N 2It keeps not
Become, needs with xN+1For mean value, greater variance and smaller weight establish a new Gaussian Profile, replace original K Gauss point
Weight is one the smallest in cloth, meanwhile, update the weights omega of K Gaussian Profilei,N,,
ωi,N+1=(1- α) ωi,N+αMi,N+1, i=1,2 ..., K.
Wherein, α is learning rate, for parameter Mi,N+1, work as xN+1When being mismatched with i-th of Gaussian Profile, Mi,N+1=1, it is no
Then, Mi,N+1=0；
(14) background estimating, according toValue it is descending, the K Gaussian Profile of block of pixels M is arranged
Sequence.B Gaussian Profile be as background model before taking, and takes rear K-B Gaussian Profile as foreground model,
Wherein, T is threshold value, determines background distributions number,
Judge whether block of pixels M belongs to moving target, as the gray value x of MN+1It is matched with some in preceding B Gaussian Profile
When, it is believed that M is background pixel；Otherwise M is object pixel；
The region of all object pixel compositions is foreground target region (vehicle target), to reduce wrong report, need to determine pass
Key range, using the bottom edge of the bounding box of foreground target as the bottom edge of key area, the width in the region is equal to the wide of vehicle target
0.8 times, a height of 60 pixel, normalizing the key area is 80x120 pixel, is denoted as Inorm。
Further, the calculating of the Haar-like feature in step (2) includes the following steps:
(21) Haar-like feature is that a kind of common feature of computer vision field describes operator, and characteristic value is one
The sum of black picture element gray value in a zonule subtracts the sum of white pixel gray value, replaces by the way of based on block feature
Calculating cost can be reduced for mode pixel-based, feature selecting major side feature, line feature, point feature (center ring around
Feature) and to corner characteristics；
(22) Haar-like feature is generally made of 2-3 rectangular block, in order to improve calculating speed, using integrogram
Method quickly calculates the sum of all gray scales in rectangle；
(23) to reduce intrinsic dimensionality, feature selecting and dimensionality reduction are carried out using PCA algorithm, the Haar- after obtaining dimensionality reduction
Like feature vector, is denoted as FHaar-like。
Further, the calculating of the co-occurrence matrix gradient orientation histogram feature in step (2) includes the following steps:
(24) normalization key area I is calculated separatelynormThe amplitude mag (x, y) of gradient and direction at position (x, y)
Ori (x, y), i.e.,
Wherein, Inorm(x, y) indicates normalization key area InormPixel value at position (x, y)；
(25) key area is divided into m × n fritter according to the ratio of width to height, fritter is non-overlapping region；
(26) it selectes a kind of offset manner to be scanned each fritter, generates a co-occurrence matrix, offset manner refers to
Relative position of the point between, totally 31 kinds, such as upper and lower relation, left-right relation etc., a kind of corresponding symbiosis of offset manner
Matrix,
The gradient direction of each pixel is carried out to pairs of combination, while original gradient direction is divided into 8, model
It encloses from 0 degree to 360 degree, i.e., every 45 degree are a bin, since every two pixel forms a combination, so co-occurrence matrix
Size is 8x8=64, and therefore, a kind of offset manner obtains the column vector of a m*n*64；
(27) it converts a kind of offset relationship to scan in entire image, until all offset manners all run-downs.In this way
The column vector of 31*64*m*n, that is, final co-occurrence matrix gradient orientation histogram feature will be generated, F is denoted asCoHOG。
Further, the calculating of local binary patterns Fourier's histogram feature LBP-HF in step (2) includes following step
It is rapid:
(28) canonical form local binary patterns are calculated using following formula,
Wherein, P indicates the quantity of pixel in neighborhood in regional area, and R indicates the radius of circle shaped neighborhood region, U (LBPP,R) indicate
Field number of pixels is P, and radius is the canonical form local binary patterns numerical value of the center pixel of the image-region of R, gcRegion
Center pixel value, gp(p=1,2 ..., P) indicates the pixel value of p-th of pixel on neighborhood；
Calculate the histogram feature of canonical form local binary patterns；
(29) discrete Fourier transform is carried out to histogram, i.e.,
Wherein, P is the number of neighbor pixel point, Up(n, r) is a certain normalized schema, hI(Up(n, r)) it is in image I
There are UpThe number of (n, r) normalized schema, H (n, u) indicate the histogram after Fourier transformation；
(210) local binary patterns Fourier histogram feature LBP-HF is obtained using following formula,
Wherein,Indicate the complex conjugate of H (n, u), FLBP- HF indicates that local binary patterns Fourier histogram is special
Sign.
Further, being modeled using continuous multiple frames of the autoregressive moving-average model to every kind of feature in step (3)
Include the following steps:
(31) autoregressive moving-average model is suitable for the analysis of objective world most of the time sequence, and unknown parameter
Few, characteristic optimizing approximation ratio is preferable, which thinks that sequence current value is present and past error and previous sequence
The linear combination of value, general type be,
ξt~WGN (0, σ2)
Wherein, p indicates that Autoregressive, q indicate sliding average order,Indicate autoregressive coefficient, θjIndicate sliding
Mean coefficient, ξtIndicate random disturbances value, ξt~WGN (0, σ2) indicate ξtIt is 0 for mean value, variance σ2Normal white noise mistake
Journey, Yi(i=1,2 ..., t.) indicates the sequential value of moment i；
(32) probability description strategy is introduced, converting above-mentioned model formation is
(33) least square problem is established, model parameter is estimated using most basic steepest descent method,
For same key area, feature contains FHaar-like, FCoHOGAnd FLBP-HFThree classes analyze settling time using sequence
Sequence signature, therefore available three kinds of different autoregressive moving-average models.
Further, the identification black smoke vehicle video clip in step (4) includes the following steps:
(41) for new vehicle target, three models are respectively used to three kinds of features that vehicle key area extracts
Classification, chooses recognition result of the recognition result as the key area of the model of maximum probability, if present frame has a pass
Key range is identified as black smoke region, then present frame is identified as black smoke frame；
(42) analysis of comprehensive multiframe, if having more than δ frame in continuous 100 frame is identified as black smoke frame, identification is worked as
Preceding video-frequency band has included black smoke vehicle, and parameter δ is the adjusting recall rate of user setting and the coefficient of rate of failing to report.
Advantages of the present invention is mainly reflected in:
(1) present invention has merged Haar-like feature, co-occurrence matrix gradient orientation histogram feature (CoHOG) and part
The advantage of binary pattern Fourier histogram feature (LBP-HF) three kinds of features, wherein Haar-like feature can be sufficiently anti-
The grey scale change situation of image is reflected, the effect of especially horizontal edge feature extraction is maximum；CoHOG feature is in illumination variation
With there is robustness, higher can the express image shape information with more refining of this feature dimension, and calculating speed under deformation
Quickly；LBP-HF not only remains traditional LBP description advantage stronger in terms of local grain description, but also increases rotation
Turn constant characteristic, while having certain robustness to shade.The advantage of comprehensive each feature, the technological invention further mention
High discrimination, reduces rate of false alarm, and have very high robustness to camera shake and vehicle shadow.
(2) analysis present invention introduces autoregressive moving-average model for the characteristic sequence of continuous multiple frames, makes full use of
The behavioral characteristics of key area feature the vehicle tail of feature and non-black smoke region that the black smoke in black smoke region is gradually spread
The feature that portion gradually translates forward and disappears, to reduce black smoke vehicle rate of false alarm.
(3) compared to conventional method, the technical scheme is that a kind of intelligence black smoke vehicle detection method, can save significantly
The manpower and financial resources for saving conventional method consumption, is conducive to the acquisition and preservation of evidence, does not influence normal traffic, can effectively mention
High law enforcement efficiency.
Detailed description of the invention
Fig. 1 is method flow schematic diagram of the invention.
Fig. 2 be the present invention use the edge feature of Haar-like feature, line feature, point feature and to corner characteristics.
Fig. 3 be the invention detects that black smoke vehicle an example.
Specific embodiment
Present invention is further described in detail with specific embodiment with reference to the accompanying drawing:
The present invention provides the black smoke vehicle recognition methods based on Gaussian Mixture and autoregressive moving-average model, can integrate difference
The static nature of type, and the behavioral characteristics that autoregressive moving-average model portrays key area are introduced, to camera shake and vehicle
Shade has very high robustness, can reduce the interference of noise, further increases the discrimination of black smoke vehicle, reduces rate of false alarm.
The present invention provides a kind of black smoke vehicle recognition methods based on Gaussian Mixture and autoregressive moving-average model, process
Figure is as shown in Figure 1, specifically follow the steps below:
Step 1: vehicle movement target is detected from traffic surveillance videos using gauss hybrid models；
Step 2: three kinds of features of extraction vehicle key area, including Haar-like feature, co-occurrence matrix gradient direction are straight
Square figure feature (CoHOG) and local binary patterns Fourier histogram feature (LBP-HF)；
Step 3: the continuous multiple frames of every kind of feature being modeled using autoregressive moving average (ARMA) model, obtain three
A difference model；
Step 4: for new vehicle target, three models being respectively used to three kinds of features that vehicle key area extracts
Classification whether there is black smoke vehicle to do current video section in conjunction with the classification results of different characteristic and the comprehensive analysis of continuous multiple frames
Judge out.
Vehicle movement target is detected from traffic surveillance videos including as follows using gauss hybrid models in the step 1
Step:
Step 1.1: converting gray level image for each frame image, and be divided into the fritter of 3x3 pixel, to each block of pixels
Gauss model is established, to improve background modeling speed and stability, reduces the interference of noise；
Step 1.2: model initialization, by taking one block of pixels M of certain frame image as an example, before the initial model of the block of pixels utilizes
The block of pixels sequence of N frame sequence image is established, and { x is denoted as1,x2,...,xN, take the average gray μ of block of pixels0And variance
The mean value and variance of the 1st Gaussian Profile are initialized, i.e.,
Wherein, I (xi,yj, k) and indicate pixel value of the block of pixels in kth frame image at position (x, y)；
Step 1.3: illustrate the update method of background model by taking N+1 frame as an example,
If the gray value x of block of pixelsN+1Meet | xN+1-μi,N|≤2.5σi,N, then the block of pixels is divided with corresponding K Gauss
Cloth matching, parameter update is as follows,
Wherein, β and ρ is the learning rate of Gaussian Profile mean value and variance respectively,Indicate Gaussian probability-density function, θ and γ
For fixed value, for the value range of regularized learning algorithm rate, xN+1For the average gray of N+1 frame block of pixels, μi,NAnd σi,N 2For
N+1 frame, the mean value and variance of i-th of Gauss model；
If the gray value x of the block of pixelsN+1It is mismatched with corresponding K Gaussian Profile, then parameter μi,NAnd σi,N 2It keeps not
Become, needs with xN+1For mean value, greater variance and smaller weight establish a new Gaussian Profile, replace original K Gauss point
Weight is one the smallest in cloth, meanwhile, update the weights omega of K Gaussian Profilei,N,,
ωi,N+1=(1- α) ωi,N+αMi,N+1, i=1,2 ..., K.
Wherein, α is learning rate, for parameter Mi,N+1, work as xN+1When being mismatched with i-th of Gaussian Profile, Mi,N+1=1, it is no
Then, Mi,N+1=0；
Step 1.4: background estimating, according toValue it is descending, to the K Gauss point of block of pixels M
Cloth sequence.B Gaussian Profile be as background model before taking, and takes rear K-B Gaussian Profile as foreground model,
Wherein, T is threshold value, determines background distributions number,
Judge whether block of pixels M belongs to moving target, as the gray value x of MN+1It is matched with some in preceding B Gaussian Profile
When, it is believed that M is background pixel；Otherwise M is object pixel；
The region of all object pixel compositions is foreground target region, to reduce wrong report, need to determine key area, in the past
The bottom edge of the bounding box of scape target is the bottom edge of key area, and wide wide 0.8 times equal to vehicle target in the region is a height of
60 pixels, normalizing the key area is 80x120 pixel, is denoted as Inorm；
The calculating of Haar-like feature in the step 2 includes the following steps:
Step 2.1:Haar-like feature is that a kind of common feature of computer vision field describes operator, characteristic value
It is that the sum of black picture element gray value in a zonule subtracts the sum of white pixel gray value, using the side based on block feature
Formula, which substitutes mode pixel-based, can reduce calculating cost, feature selecting major side feature, line feature, point feature (center
Ring characteristics) and to corner characteristics, as shown in Fig. 2, every a line represents a kind of feature；
Step 2.2:Haar-like feature is generally made of 2-3 rectangular block, in order to improve calculating speed, using integral
The method of figure quickly calculates the sum of all gray scales in rectangle；
Step 2.3: to reduce intrinsic dimensionality, feature selecting and dimensionality reduction are carried out using PCA algorithm, after obtaining dimensionality reduction
Haar-like feature vector, is denoted as FHaar-like。
The calculating of co-occurrence matrix gradient orientation histogram feature in the step 2 includes the following steps:
Step 2.4: calculating separately normalization key area InormThe amplitude mag (x, y) of gradient and side at position (x, y)
To ori (x, y), i.e.,
Wherein, Inorm(x, y) indicates normalization key area InormPixel value at position (x, y)；
Step 2.5: key area being divided into m × n fritter according to the ratio of width to height, fritter is non-overlapping region；
Step 2.6: selecting a kind of offset manner and each fritter is scanned, generate a co-occurrence matrix, offset manner
It is the relative position given directions between, totally 31 kinds, such as upper and lower relation, left-right relation etc., a kind of offset manner correspondence one
Co-occurrence matrix,
The gradient direction of each pixel is carried out to pairs of combination, while original gradient direction is divided into 8, model
It encloses from 0 degree to 360 degree, i.e., every 45 degree are a bin, since every two pixel forms a combination, so co-occurrence matrix
Size is 8x8=64, and therefore, a kind of offset manner obtains the column vector of a m*n*64；
Step 2.7: converting a kind of offset relationship and scanned in entire image, until all offset manners all run-downs.
The column vector of 31*64*m*n, that is, final co-occurrence matrix gradient orientation histogram feature will be generated in this way, be denoted as
FCoHOG。
The calculating of local binary patterns Fourier's histogram feature in the step 2 includes the following steps:
Step 2.8: canonical form local binary patterns are calculated using following formula,
Wherein, P indicates the quantity of pixel in neighborhood in regional area, and R indicates the radius of circle shaped neighborhood region, U (LBPP,R) indicate
Field number of pixels is P, and radius is the canonical form local binary patterns numerical value of the center pixel of the image-region of R, gcRegion
Center pixel value, gp(p=1,2 ..., P) indicates the pixel value of p-th of pixel on neighborhood；
Calculate the histogram feature of canonical form local binary patterns；
Step 2.9: discrete Fourier transform being carried out to histogram, i.e.,
Wherein, P is the number of neighbor pixel point, Up(n, r) is a certain normalized schema, hI(Up(n, r)) it is in image I
There are UpThe number of (n, r) normalized schema, H (n, u) indicate the histogram after Fourier transformation；
Step 2.10: local binary patterns Fourier's histogram feature is obtained using following formula,
Wherein,Indicate the complex conjugate of H (n, u), FLBP- HF indicates that local binary patterns Fourier histogram is special
Sign.
Modeling is carried out including such as using continuous multiple frames of the autoregressive moving-average model to every kind of feature in the step 3
Lower step:
Step 3.1: autoregressive moving-average model is suitable for the analysis of objective world most of the time sequence, and unknown
Parameter is few, and characteristic optimizing approximation ratio is preferable, the model think sequence current value be present and past error and previously
The linear combination of sequential value, general type be,
ξt~WGN (0, σ2)
Wherein, p indicates that Autoregressive, q indicate sliding average order,Indicate autoregressive coefficient, θjIt indicates to slide
Dynamic mean coefficient, ξtIndicate random disturbances value, ξt~WGN (0, σ2) indicate ξtIt is 0 for mean value, variance σ2Normal white noise
Process, Yi(i=1,2 ..., t.) indicates the sequential value of moment i；
Step 3.2: introducing probability description strategy, converting above-mentioned model formation is
Step 3.3: least square problem is established, model parameter is estimated using most basic steepest descent method,
For same key area, feature contains FHaar-like, FCoHOGAnd FLBP-HFThree classes analyze settling time using sequence
Sequence signature, therefore available three kinds of different autoregressive moving-average models.
Identification black smoke vehicle video clip in the step 4 includes the following steps:
Step 4.1: for new vehicle target, three models being respectively used to three kinds of spies that vehicle key area extracts
The classification of sign chooses recognition result of the recognition result as the key area of the model of maximum probability, if present frame has one
A key area is identified as black smoke region, then present frame is identified as black smoke frame；
Step 4.2: the analysis of comprehensive multiframe is recognized if having more than δ frame in continuous 100 frame is identified as black smoke frame
Settled preceding video-frequency band has included black smoke vehicle；
Parameter δ is the adjusting recall rate of user setting and the coefficient of rate of failing to report.
Fig. 3 shows the example of the black smoke vehicle detected from vehicle monitoring video using the present invention, filament black
Rectangle frame indicates the band of position of the foreground target detected, and thick line black rectangle frame indicates the position of key area, rectangle frame
The numerical value of bottom indicates that the region is the probability in black smoke region.
The above described is only a preferred embodiment of the present invention, being not the limit for making any other form to the present invention
System, and made any modification or equivalent variations according to the technical essence of the invention, still fall within present invention model claimed
It encloses.
Claims (7)
1. the black smoke vehicle recognition methods based on Gaussian Mixture and autoregressive moving-average model, specific step is as follows, and feature exists
In:
(1) vehicle movement target is detected from traffic surveillance videos using gauss hybrid models；
(2) three kinds of features of vehicle key area, including Haar-like feature are extracted, co-occurrence matrix gradient orientation histogram is special
It seeks peace local binary patterns Fourier's histogram feature；
(3) it is modeled using continuous multiple frames of the autoregressive moving-average model to every kind of feature, obtains three different models；
(4) for new vehicle target, three models are respectively used to the classification for three kinds of features that vehicle key area extracts,
In conjunction with the classification results of different characteristic and the comprehensive analysis of continuous multiple frames, whether there is black smoke vehicle to judge current video section.
2. the black smoke vehicle recognition methods according to claim 1 based on Gaussian Mixture and autoregressive moving-average model,
It is characterized in that: detecting vehicle movement target from traffic surveillance videos including as follows using gauss hybrid models in step (1)
Step:
(11) gray level image is converted by each frame image, and is divided into the fritter of 3x3 pixel, Gauss is established to each block of pixels
Model reduces the interference of noise to improve background modeling speed and stability；
(12) model initialization, by taking one block of pixels M of certain frame image as an example, the initial model of the block of pixels utilizes preceding N frame sequence
The block of pixels sequence of image is established, and { x is denoted as1,x2,...,xN, take the average gray μ of block of pixels0And varianceInitialization the
The mean value and variance of 1 Gaussian Profile, i.e.,
Wherein, I (xi,yj, k) and indicate pixel value of the block of pixels in kth frame image at position (x, y)；
(13) illustrate the update method of background model by taking N+1 frame as an example,
If the gray value x of block of pixelsN+1Meet | xN+1-μi,N|≤2.5σi,N, then the block of pixels and corresponding K Gaussian Profile
To match, parameter update is as follows,
Wherein, β and ρ is the learning rate of Gaussian Profile mean value and variance respectively,Indicate Gaussian probability-density function, θ and γ are solid
Definite value, for the value range of regularized learning algorithm rate, xN+1For the average gray of N+1 frame block of pixels, μi,NAnd σi,N 2For N+1
Frame, the mean value and variance of i-th of Gauss model；
If the gray value x of the block of pixelsN+1It is mismatched with corresponding K Gaussian Profile, then parameter μi,NAnd σi,N 2It remains unchanged, needs
It will be with xN+1For mean value, greater variance and smaller weight establish a new Gaussian Profile, replace and weigh in original K Gaussian Profile
Weight is one the smallest, meanwhile, update the weights omega of K Gaussian Profilei,N,,
ωi,N+1=(1- α) ωi,N+αMi,N+1, i=1,2 ..., K.
Wherein, α is learning rate, for parameter Mi,N+1, work as xN+1When being mismatched with i-th of Gaussian Profile, Mi,N+1=1, otherwise,
Mi,N+1=0；
(14) background estimating, according toValue it is descending, sort to the K Gaussian Profile of block of pixels M.It takes
Preceding B Gaussian Profile takes rear K-B Gaussian Profile as foreground model as background model,
Wherein, T is threshold value, determines background distributions number,
Judge whether block of pixels M belongs to moving target, as the gray value x of MN+1When being matched with some in preceding B Gaussian Profile, recognize
It is background pixel for M；Otherwise M is object pixel；
The region of all object pixel compositions is foreground target region, to reduce wrong report, key area need to be determined, with prospect mesh
The bottom edge of target bounding box is the bottom edge of key area, wide wide 0.8 times equal to vehicle target in the region, a height of 60 picture
Element, normalizing the key area is 80x120 pixel, is denoted as Inorm。
3. the black smoke vehicle recognition methods according to claim 1 based on Gaussian Mixture and autoregressive moving-average model,
Be characterized in that: the calculating of the Haar-like feature in step (2) includes the following steps:
(21) Haar-like feature is that a kind of common feature of computer vision field describes operator, and characteristic value is one small
The sum of black picture element gray value in region subtracts the sum of white pixel gray value, substitutes and is based on by the way of based on block feature
The mode of pixel can reduce calculating cost, feature selecting major side feature, line feature, point feature (center ring characteristics) and
To corner characteristics；
(22) Haar-like feature is generally made of 2-3 rectangular block, in order to improve calculating speed, using the method for integrogram
Quickly calculate the sum of all gray scales in rectangle；
(23) to reduce intrinsic dimensionality, feature selecting and dimensionality reduction are carried out using PCA algorithm, the Haar-like after obtaining dimensionality reduction is special
Vector is levied, F is denoted asHaar-like。
4. the black smoke vehicle recognition methods based on Gaussian Mixture and autoregressive moving-average model as described in claim 1, special
Sign is that the calculating of the co-occurrence matrix gradient orientation histogram feature in step (2) includes the following steps:
(24) normalization key area I is calculated separatelynormAt position (x, y) gradient amplitude mag (x, y) and direction ori (x,
Y), i.e.,
Wherein, Inorm(x, y) indicates normalization key area InormPixel value at position (x, y)；
(25) key area is divided into m × n fritter according to the ratio of width to height, fritter is non-overlapping region；
(26) it selectes a kind of offset manner to be scanned each fritter, generates a co-occurrence matrix, offset manner is to give directions pair
Between relative position, totally 31 kinds, such as upper and lower relation, left-right relation etc., a kind of corresponding co-occurrence matrix of offset manner,
The gradient direction of each pixel carries out to pairs of combination, while original gradient direction is divided into 8, range from
0 degree to 360 degree, i.e., every 45 degree are a bin, since every two pixel forms a combination, so the size of co-occurrence matrix
For 8x8=64, therefore, a kind of offset manner obtains the column vector of a m*n*64；
(27) it converts a kind of offset relationship to scan in entire image, until all offset manners all run-downs.In this way will
The column vector of 31*64*m*n, that is, final co-occurrence matrix gradient orientation histogram feature are generated, F is denoted asCoHOG。
5. the black smoke vehicle recognition methods according to claim 1 based on Gaussian Mixture and autoregressive moving-average model,
Be characterized in that: the calculating of local binary patterns Fourier's histogram feature in step (2) includes the following steps:
(28) canonical form local binary patterns are calculated using following formula,
Wherein, P indicates the quantity of pixel in neighborhood in regional area, and R indicates the radius of circle shaped neighborhood region, U (LBPP,R) indicate field
Number of pixels is P, and radius is the canonical form local binary patterns numerical value of the center pixel of the image-region of R, gcThe center in region
Pixel value, gp(p=1,2 ..., P) indicates the pixel value of p-th of pixel on neighborhood；
Calculate the histogram feature of canonical form local binary patterns；
(29) discrete Fourier transform is carried out to histogram, i.e.,
Wherein, P is the number of neighbor pixel point, Up(n, r) is a certain normalized schema, hI(Up(n, r)) it is that there are U in image Ip
The number of (n, r) normalized schema, H (n, u) indicate the histogram after Fourier transformation；
(210) local binary patterns Fourier's histogram feature is obtained using following formula,
Wherein,Indicate the complex conjugate of H (n, u), FLBP-HFIndicate local binary patterns Fourier histogram feature.
6. the black smoke vehicle recognition methods according to claim 1 based on Gaussian Mixture and autoregressive moving-average model,
It is characterized in that: carrying out modeling including such as using continuous multiple frames of the autoregressive moving-average model to every kind of feature in step (3)
Lower step:
(31) autoregressive moving-average model is suitable for the analysis of objective world most of the time sequence, and unknown parameter is few,
Characteristic optimizing approximation ratio is preferable, which thinks that sequence current value is present and past error and previous sequential value
Linear combination, general type be,
ξt~WGN (0, σ2)
Wherein, p indicates that Autoregressive, q indicate sliding average order,Indicate autoregressive coefficient, θjIndicate sliding average
Coefficient, ξtIndicate random disturbances value, ξ t~WGN (0, σ2) indicate ξtIt is 0 for mean value, variance σ2Normal white noise process, Yi
(i=1,2 ..., t.) indicates the sequential value of moment i；
(32) probability description strategy is introduced, converting above-mentioned model formation is
(33) least square problem is established, model parameter is estimated using most basic steepest descent method,
For same key area, feature contains FHaar-like, FCoHOGAnd FLBP-HFThree classes analyze settling time sequence using sequence
Feature, therefore available three kinds of different autoregressive moving-average models.
7. the black smoke vehicle recognition methods according to claim 1 based on Gaussian Mixture and autoregressive moving-average model,
Be characterized in that: the identification black smoke vehicle video clip in step (4) includes the following steps:
(41) for new vehicle target, three models are respectively used to point for three kinds of features that vehicle key area extracts
Class chooses recognition result of the recognition result as the key area of the model of maximum probability, if present frame has a key
Region is identified as black smoke region, then present frame is identified as black smoke frame；
(42) analysis of comprehensive multiframe, if having more than δ frame in continuous 100 frame is identified as black smoke frame, forward sight is worked as in identification
Contain black smoke vehicle in frequency range, parameter δ is the adjusting recall rate of user setting and the coefficient of rate of failing to report.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810789111.3A CN109190455B (en) | 2018-07-18 | 2018-07-18 | Black smoke vehicle identification method based on Gaussian mixture and autoregressive moving average model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810789111.3A CN109190455B (en) | 2018-07-18 | 2018-07-18 | Black smoke vehicle identification method based on Gaussian mixture and autoregressive moving average model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109190455A true CN109190455A (en) | 2019-01-11 |
| CN109190455B CN109190455B (en) | 2021-08-13 |
Family
ID=64936156
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810789111.3A Active CN109190455B (en) | 2018-07-18 | 2018-07-18 | Black smoke vehicle identification method based on Gaussian mixture and autoregressive moving average model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109190455B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110084220A (en) * | 2019-05-08 | 2019-08-02 | 重庆邮电大学 | A kind of vehicle-mounted fatigue detection method based on multiple dimensioned binary mode |
| CN111523492A (en) * | 2020-04-26 | 2020-08-11 | 安徽皖仪科技股份有限公司 | Detection method of black smoke vehicle |
| CN111539343A (en) * | 2020-04-26 | 2020-08-14 | 安徽皖仪科技股份有限公司 | Black smoke vehicle detection method based on convolution attention network |
| CN112598648A (en) * | 2020-12-24 | 2021-04-02 | 重庆邮电大学 | Image seam cutting tampering detection method based on image gradient direction |
| CN117788570A (en) * | 2024-02-26 | 2024-03-29 | 山东济矿鲁能煤电股份有限公司阳城煤矿 | Bucket wheel machine positioning method and system based on machine vision |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120089545A1 (en) * | 2009-04-01 | 2012-04-12 | Sony Corporation | Device and method for multiclass object detection |
| CN103714554A (en) * | 2013-12-12 | 2014-04-09 | 华中科技大学 | Video tracking method based on spread fusion |
| CN104680508A (en) * | 2013-11-29 | 2015-06-03 | 华为技术有限公司 | Convolutional neural network and target object detection method based on convolutional neural network |
| CN105703954A (en) * | 2016-03-17 | 2016-06-22 | 福州大学 | Network data flow prediction method based on ARIMA model |
| CN106780548A (en) * | 2016-11-16 | 2017-05-31 | 南宁市浩发科技有限公司 | moving vehicle detection method based on traffic video |
-
2018
- 2018-07-18 CN CN201810789111.3A patent/CN109190455B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120089545A1 (en) * | 2009-04-01 | 2012-04-12 | Sony Corporation | Device and method for multiclass object detection |
| CN104680508A (en) * | 2013-11-29 | 2015-06-03 | 华为技术有限公司 | Convolutional neural network and target object detection method based on convolutional neural network |
| CN103714554A (en) * | 2013-12-12 | 2014-04-09 | 华中科技大学 | Video tracking method based on spread fusion |
| CN105703954A (en) * | 2016-03-17 | 2016-06-22 | 福州大学 | Network data flow prediction method based on ARIMA model |
| CN106780548A (en) * | 2016-11-16 | 2017-05-31 | 南宁市浩发科技有限公司 | moving vehicle detection method based on traffic video |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110084220A (en) * | 2019-05-08 | 2019-08-02 | 重庆邮电大学 | A kind of vehicle-mounted fatigue detection method based on multiple dimensioned binary mode |
| CN111523492A (en) * | 2020-04-26 | 2020-08-11 | 安徽皖仪科技股份有限公司 | Detection method of black smoke vehicle |
| CN111539343A (en) * | 2020-04-26 | 2020-08-14 | 安徽皖仪科技股份有限公司 | Black smoke vehicle detection method based on convolution attention network |
| CN111523492B (en) * | 2020-04-26 | 2023-04-18 | 安徽皖仪科技股份有限公司 | Detection method of black smoke vehicle |
| CN111539343B (en) * | 2020-04-26 | 2023-05-23 | 安徽皖仪科技股份有限公司 | Black smoke vehicle detection method based on convolution attention network |
| CN112598648A (en) * | 2020-12-24 | 2021-04-02 | 重庆邮电大学 | Image seam cutting tampering detection method based on image gradient direction |
| CN117788570A (en) * | 2024-02-26 | 2024-03-29 | 山东济矿鲁能煤电股份有限公司阳城煤矿 | Bucket wheel machine positioning method and system based on machine vision |
| CN117788570B (en) * | 2024-02-26 | 2024-05-07 | 山东济矿鲁能煤电股份有限公司阳城煤矿 | Bucket wheel machine positioning method and system based on machine vision |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109190455B (en) | 2021-08-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109816024B (en) | Real-time vehicle logo detection method based on multi-scale feature fusion and DCNN | |
| CN109190455A (en) | Black smoke vehicle recognition methods based on Gaussian Mixture and autoregressive moving-average model | |
| CN102165493B (en) | Detection of vehicles in an image | |
| CN107610114B (en) | optical satellite remote sensing image cloud and snow fog detection method based on support vector machine | |
| CN104778444B (en) | The appearance features analysis method of vehicle image under road scene | |
| CN103605953B (en) | Vehicle interest target detection method based on sliding window search | |
| CN109657632B (en) | Lane line detection and identification method | |
| CN103218831B (en) | A kind of video frequency motion target classifying identification method based on profile constraint | |
| CN109284669A (en) | Pedestrian detection method based on Mask RCNN | |
| CN105447503B (en) | Pedestrian detection method based on rarefaction representation LBP and HOG fusion | |
| CN107103317A (en) | Fuzzy license plate image recognition algorithm based on image co-registration and blind deconvolution | |
| CN109255326B (en) | Traffic scene smoke intelligent detection method based on multi-dimensional information feature fusion | |
| CN106650553A (en) | License plate recognition method and system | |
| CN109101924A (en) | A kind of pavement marking recognition methods based on machine learning | |
| CN109918971B (en) | Method and device for detecting number of people in monitoring video | |
| CN106529532A (en) | License plate identification system based on integral feature channels and gray projection | |
| CN106529461A (en) | Vehicle model identifying algorithm based on integral characteristic channel and SVM training device | |
| Li et al. | Research on vehicle license plate location based on neural networks | |
| CN110766016B (en) | Code-spraying character recognition method based on probabilistic neural network | |
| CN109191495B (en) | Black smoke vehicle detection method based on self-organizing background difference model and multi-feature fusion | |
| CN105718866A (en) | Visual target detection and identification method | |
| CN105224945B (en) | A kind of automobile logo identification method based on joint-detection and identification algorithm | |
| CN106503748A (en) | A kind of based on S SIFT features and the vehicle targets of SVM training aids | |
| CN113034378B (en) | Method for distinguishing electric automobile from fuel automobile | |
| CN110706235A (en) | Far infrared pedestrian detection method based on two-stage cascade segmentation |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |