CN109241824B - Intelligent black smoke vehicle monitoring method based on codebook and smooth conversion autoregressive model - Google Patents
Intelligent black smoke vehicle monitoring method based on codebook and smooth conversion autoregressive model Download PDFInfo
- Publication number
- CN109241824B CN109241824B CN201810786681.7A CN201810786681A CN109241824B CN 109241824 B CN109241824 B CN 109241824B CN 201810786681 A CN201810786681 A CN 201810786681A CN 109241824 B CN109241824 B CN 109241824B
- Authority
- CN
- China
- Prior art keywords
- model
- codebook
- value
- histogram
- code word
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
- G06V20/42—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items of sport video content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a black smoke vehicle intelligent monitoring method based on a codebook and a smooth conversion autoregressive model, which comprises the following steps: (1) detecting a moving target in a video by using a codebook model, and determining a key area behind the target; (2) extracting Tamura characteristics, central symmetry local binary pattern histogram characteristics and gray level histogram characteristics of the key area, and combining to form three types of static characteristics of the key area; (3) respectively modeling the time sequences of the three types of static characteristics by using a smooth transition autoregressive model, and taking the solution of the model as a final characteristic vector for depicting the dynamic characteristics of a key area; (4) and training the three types of characteristics to obtain three SVM classifiers, performing weighted fusion on the three classification results to obtain a recognition result of the current key area, and judging whether the current video segment has black smoke cars or not through analysis of the recognition results of a plurality of continuous key areas. The method and the device can improve the detection efficiency of the black smoke vehicle in the road monitoring video and reduce the false alarm rate.
Description
Technical Field
The invention relates to the technical field of image feature extraction and time sequence analysis, in particular to an intelligent black smoke vehicle monitoring method based on a codebook and a smooth conversion self-regression model.
Background
The pollution of the diesel truck is always the most important of the pollution of the motor vehicle, and the black smoke vehicle discussed by the invention is a common diesel truck, so that how to find the black smoke vehicle running on the road in time and implement elimination or forced maintenance according to the pollution degree is very helpful for reducing air pollution, fighting against rigor and fighting against pollution of the diesel truck is well achieved, the pollutant emission is greatly reduced, and the victory of the defense war in blue sky is won.
At present, vehicle tail gas analysis devices are adopted in partial areas to detect black smoke vehicles, but the black smoke vehicles have many defects, on one hand, the vehicle tail gas analysis devices are often expensive, and a large amount of funds are consumed for subsequent maintenance, and the feasibility of configuring each vehicle is difficult due to the increase of the number of the vehicles; on the other hand, for the analysis of the exhaust gas of a roadside-mounted vehicle, the detection result is often influenced by various factors, such as multiple vehicles running in parallel, too close to the vehicle, bad weather (wind speed is too high, rain and snow weather), road environment background, vehicle emission beside, height of an exhaust pipe, equipment conditions (light intensity, noise and the like), professional level of an operator (installation and debugging) and the like.
Disclosure of Invention
The invention aims to solve the technical problem of providing an intelligent black smoke vehicle monitoring method based on a codebook and a smooth conversion autoregressive model, which can further improve the recognition rate of black smoke vehicle detection and reduce the false alarm rate.
In order to solve the technical problem, the invention provides an intelligent black smoke vehicle monitoring method based on a codebook model and a smooth conversion autoregressive model, which comprises the following steps:
(1) detecting a moving target in a video by using a Codebook (Codebook) model, and determining a key area behind the target;
(2) extracting Tamura characteristics, central symmetry local binary pattern (CS-LBP) histogram characteristics and gray level histogram characteristics of a key area, and combining to form three types of static characteristics of the key area;
(3) respectively modeling the time sequences of the three types of static features by using a Smooth Transition Autoregressive (STAR) model, and taking the solution of the model as a final feature vector for depicting the dynamic features of a key region;
(4) and training the three types of characteristics to obtain three SVM classifiers, performing weighted fusion on the three classification results to obtain a recognition result of the current key area, and judging whether the current video segment has a black smoke car or not through analysis of the recognition results of a plurality of continuous key areas. The method provides a black smoke vehicle intelligent monitoring method based on a codebook and a smooth conversion autoregressive model, the method detects a moving target by utilizing the codebook model, integrates Tamura characteristics, central symmetry local binary pattern (CS-LBP) histogram characteristics and gray level histogram characteristics, and is combined with the smooth conversion autoregressive model to depict the time sequence characteristics of a key area of a vehicle, and the method can greatly improve the efficiency of monitoring the black smoke vehicle.
As an improvement of the present invention, the detection of the moving object in the video by using a Codebook (Codebook) model in step (1) includes the following points:
(11) codebook model initialization and model characterization
Let C ═ C1,c2,...cL) For a one-pixel codebook, the codebook for each pixel is set to null, i.e., L is 0, where L represents the number of codewords in each codebook, and each codeword c is set to zeroi(i 1, 2.. L.) is represented by a gray value viAnd hexahydric group auxiThe two parts are formed into a whole body,
wherein the content of the first and second substances,andrespectively representing the minimum and maximum grey values, f, in the code wordiIndicating the frequency of occurrence of the code word, λiIndicating the maximum time interval, p, during which the codeword does not occur during trainingiAnd q isiRespectively representing the time of the first and last occurrence of the codeword;
(12) codebook modeling and model updating
i) Let X be { X ═ X1,x2,...,xNIs a sequence of sample values for a pixel, xtIs the gray value at the moment t of the current pixel, for each x in the sampled value of each pixeltFinding out the code word c matched with it by using the following conditionm,
Wherein [ I ]low,Ihigh]Representing the range of variation of the grey value of each code word, epsilon1Is a threshold value, vmRepresents a code word cmCorresponding gray values;
ii) if the codebook is empty or no matching codeword is found, a new codeword is created,
if there is a code word cmThe code word is updated to be the code word when the two conditions are met
iii) after the training is over, calculating the maximum time interval during which each code word of the pixel does not reappear, and for each code word ci(i=1,2,...,L.),λi=max{λi,(N-qi+pi-1) using λ to eliminate redundant code words, and obtaining an initial codebook M ═ c that best represents the real backgroundk|ck∈C,λk≤Tm},TmUsually half the number of training frames;
(13) foreground target extraction
For a new input pixel xtIf a matching codeword c satisfying the following condition can be found in the codebook MmThen xtThe foreground moving object pixels, otherwise the background pixels,
wherein [ I ]low,Ihigh]Representing the range of variation of the grey value of each code word, epsilon2Is a threshold value, vmRepresents a code word cmThe corresponding gray value;
as a modification of the present invention, the Tamura characteristics of the key regions in step (2) include the following:
(21) the method is characterized in that: the roughness (COA) is calculated by,
where m and n denote the width and height of the image, { E1,E2,...,Ek,...,ELDenotes the horizontal average intensity difference E at position (i, j) when k takes values in the set {1,2,3,4, 5} throughk,h(i, j) and the difference E in vertical average intensityk,v(ii) a set of values of (i, j),
Ek,v(i,j)=|Ak(i,j+2k-1)-Ak(i,j-2k-1)|
Ek,h(i,j)=|Ak(i+2k-1,j)-Ak(i-2k-1,j)|
wherein F (i, j) represents the grayscale value at position (i, j), Ak(x1,x2) Indicates that the active window is 2k×2k(k is the average value of 1,2,3,4, 5) pixels.
(22) The second characteristic: a contrast ratio (CON) is calculated by,
wherein σ represents the standard deviation, μ4Fourth moment of mean, parameter n0Typically set to 0.25.
(23) The characteristics are three: a direction Degree (DIR) calculated by,
where φ represents the quantization direction code, p represents the histogram peak, and for any one peak p, WpRepresenting the quantization range contained by the peak, npIndicates the number of peaks in the histogram, phipRepresents WpMiddle maximumQuantized values in histograms, HDRepresenting a histogram, by calculating the number of points at which the gradient magnitude | Δ G | is greater than the threshold tho and the angle θ is within a certain range, HD(k) The number of points representing the point amplitude | Δ G | ≧ tho and the angle θ satisfying (2k-1) π/2n ≦ θ ≦ (2k +1) π/2n, θ and | Δ G |, are calculated by
θ=tan-1(ΔV/ΔH)+π/2
|ΔG|=(|ΔH|+|ΔV|)/2
Wherein, DeltaVAnd ΔHObtaining a vertical difference value and a horizontal difference value through a 3 multiplied by 3 Prewitt operator;
(24) the characteristics are as follows: the Linearity (LIN) is calculated by,
wherein, PDd(i, j) denotes the value of the element of the n × n local direction co-occurrence matrix (DCOM) at a point (i, j) which is a distance d and has a direction code i and a direction code j along the boundary, we calculate H using differentiation and thresholdingDAnd PDdThe local direction co-occurrence matrix is defined as the relative frequency of two adjacent units, and is separated by a convenient distance d along the image, wherein the value is +1 when the local direction co-occurrence matrix is the same as the original direction, and the value is-1 when the local direction co-occurrence matrix is vertical to the original direction;
(25) the characteristics are as follows: the Regularity (REG) is calculated by,
REG=1-r(σCON+σCOR+σDIR+σLIN)
wherein σXXXR is a normalization factor corresponding to the standard deviation of feature XXX;
(26) the characteristics are as follows: the rough degree (ROU) is calculated by,
ROU=COA+CON
wherein COA and CON represent roughness and contrast, respectively;
the first type of characteristic Tamura characteristic is denoted as FTamuraI.e. FTamura={CON,COR,DIR,LIN,REG,ROU}。
The calculation of the feature of the central symmetry local binary pattern (CS-LBP) histogram of the key region in the step (2) comprises the following steps:
(27) the centrosymmetric local binary pattern is calculated in such a way,
wherein R represents the radius, N represents the number of pixels captured in the neighborhood, T represents the threshold, CSLBPR,N,T(x, y) represents the value of the centrosymmetric local binary pattern feature with radius of R and neighborhood point number of N at the position (x, y);
(28) extracting histogram features from the obtained centrosymmetric local binary pattern map;
let feature of the second kind be denoted as FCSLBPI.e., (28) extracted histogram features.
The gray histogram feature of the key area in the step (2) comprises the following steps:
(29) converting the key area image into a gray image, and performing a gray value cutting process, namely, if the key area image is larger than the threshold value ThAll gray values of (1) are assigned as Ttru,
Wherein, Inorm(x, y) represents the grayscale value of the normalized key region at position (x, y);
(210) calculating the gray level histogram of the key region image after the truncation processing, wherein the gray level distribution is [0, Ttru]Wherein T istruGenerally taking 180;
let feature type three be denoted as FHistI.e., (29) extracted histogram features.
As an improvement of the present invention, the modeling the time series of the static features by using the STAR model in the step (3) includes the following steps:
(31) the STAR model can be expressed as
Wherein, ai(i ═ 1,2,. cndot., P) and bi(i 1, 2.. P.) are all parameters to be estimated, ytIs the time series to be investigated, a0And b0Is a constant, P denotes the order of the investigated sequence, εtIs independent and equally distributed random variable, G (-) is a continuous smooth function which is divided into a logistic function and an exponential function,
G(yt-d;γ,c)=[1+exp(-γ×(yt-d-c))]-1;γ>0
G(yt-d;γ,c)=1-exp(-γ×(yt-d-c)2);γ>0
wherein, yt-dIs a threshold variable, d is a delay parameter, c is a threshold value between the two mechanisms;
(32) determining a parameter P according to experience, and performing nonlinear least square estimation on the model to obtain an estimated value of the model parameter;
(33) respectively modeling the time sequences of the three types of static features by using a STAR model, and taking the solution of the model as a final feature vector for depicting the dynamic features of the key region
As an improvement of the present invention, the determination of whether there is a black smoke car in the current video segment in step (4) includes the following steps:
(41) obtaining different model parameters by using different training samples, training an SVM model by using the parameters, and obtaining three SVM classifiers by using three types of feature training for classifying new estimation parameters;
(42) respectively using the three models for classification of new features, and performing weighted fusion on classification results to obtain an identification result of the current key region;
wherein S iskRepresenting the classification result of the kth feature classifier;
(43) and judging whether the current video segment has black smoke cars or not through the analysis of the identification results of a plurality of continuous key areas.
The invention has the beneficial effects that: (1) the codebook model is adopted to model the background, the moving target is detected, the anti-interference capability is strong, the false detection rate is low, the calculation complexity is small, and the codebook model is very suitable for real-time detection; (2) the technology integrates Tamura characteristics, central symmetry local binary pattern (CS-LBP) histogram characteristics and gray level histogram characteristics. The Tamura characteristic is provided according to human psychology research on texture visual perception, high-level visual characteristics of textures can be captured well, gradient information is blended into a centrosymmetric local binary pattern, and compared with a traditional local binary pattern, the characteristic of a key area can be better described; (3) the method combines a smooth conversion autoregressive model to depict the time sequence characteristics of the key area of the vehicle, and the fusion of multiple characteristics and the utilization of the time sequence characteristics greatly improve the detection rate and reduce the false alarm rate;
drawings
FIG. 1 is a schematic flow chart of the method of the present invention.
Fig. 2 shows an example of a black smoke car detected by the present invention.
Detailed Description
The invention provides a black smoke vehicle intelligent monitoring method based on a codebook and a smooth conversion autoregressive model, a flow chart of which is shown in figure 1 and is specifically carried out according to the following steps:
step 1: detecting a moving target in a video by using a Codebook (Codebook) model, and determining a key area behind the target;
step 2: extracting Tamura characteristics, central symmetry local binary pattern (CS-LBP) histogram characteristics and gray level histogram characteristics of a key area, and combining to form three types of static characteristics of the key area;
and step 3: modeling the time sequences of the three types of static features by using a Smooth Transition Autoregressive (STAR) model, and taking the solution of the model as a final feature vector for depicting the dynamic features of a key region;
and 4, step 4: and training the three types of characteristics to obtain three SVM classifiers, performing weighted fusion on the three classification results to obtain a recognition result of the current key area, and judging whether the current video segment has black smoke cars or not through analysis of the recognition results of a plurality of continuous key areas.
The detection of the moving object in the video by using a Codebook (Codebook) model in the step 1 includes the following points:
step 1.1: codebook model initialization and model characterization
Let C ═ C1,c2,...cL) For a one-pixel codebook, the codebook for each pixel is set to null, i.e., L is 0, where L represents the number of codewords in each codebook, and each codeword c is set to zeroi(i 1, 2.. L.) is represented by a gray value viAnd hexahydric group auxiThe two parts are formed into a whole body,
wherein the content of the first and second substances,andrespectively representing the minimum and maximum grey values, f, in the code wordiIndicating the frequency of occurrence of the code word, λiIndicating the maximum time interval, p, during which the codeword does not occur during trainingiAnd q isiRespectively representing the time of the first and last occurrence of the codeword;
step 1.2: codebook modeling and model updating
i) Let X be { X ═ X1,x2,...,xNIs a sequence of sample values for a pixel, xtIs the gray value at the moment t of the current pixel, for each x in the sampled value of each pixeltFinding out the code word c matched with it by using the following conditionm,
Wherein [ I ]low,Ihigh]Representing the range of variation of the grey value of each code word, epsilon1Is a threshold value, vmRepresents a code word cmCorresponding gray values;
ii) if the codebook is empty or no matching codeword is found, a new codeword is created,
if there is a code word cmThe code word is updated to be the code word when the two conditions are met
iii) after the training is over, calculating the maximum time interval during which each code word of the pixel does not reappear, and for each code word ci(i=1,2,...,L.),λi=max{λi,(N-qi+pi-1) using λ to eliminate redundant code words, and obtaining an initial codebook M ═ c that best represents the real backgroundk|ck∈C,λk≤Tm},TmUsually half the number of training frames;
step 1.3: foreground target extraction
For a new input pixel xtIf a matching codeword c satisfying the following condition can be found in the codebook MmThen xtThe foreground moving object pixels, otherwise the background pixels,
wherein [ I ]low,Ihigh]Representing the range of variation of the grey value of each code word, epsilon2Is a threshold value, vmRepresents a code word cmThe corresponding gray value;
the Tamura characteristics of the key regions in step 2 include the following:
step 2.1: the method is characterized in that: the roughness (COA) is calculated by,
where m and n denote the width and height of the image, { E1,E2,...,Ek,...,ELDenotes the horizontal average intensity difference E at position (i, j) when k takes values in the set {1,2,3,4, 5} throughk,h(i, j) and the difference E in vertical average intensityk,v(ii) a set of values of (i, j),
Ek,v(i,j)=|Ak(i,j+2k-1)-Ak(i,j-2k-1)|
Ek,h(i,j)=|Ak(i+2k-1,j)-Ak(i-2k-1,j)|
wherein F (i, j) represents the grayscale value at position (i, j), Ak(x1,x2) Indicates that the active window is 2k×2k(k is the average value of 1,2,3,4, 5) pixels.
Step 2.2: the second characteristic: a contrast ratio (CON) is calculated by,
wherein σ represents the standard deviation, μ4Fourth moment of mean, parameter n0Typically set to 0.25.
Step 2.3: the characteristics are three: a direction Degree (DIR) calculated by,
where φ represents the quantization direction code, p represents the histogram peak, and for any one peak p, WpRepresenting the quantization range contained by the peak, npIndicates the number of peaks in the histogram, phipRepresents WpQuantized values in the medium-maximum histogram, HDRepresenting a histogram, by calculating the number of points at which the gradient magnitude | Δ G | is greater than the threshold tho and the angle θ is within a certain range, HD(k) The number of points representing the point amplitude | Δ G | ≧ tho and the angle θ satisfying (2k-1) π/2n ≦ θ ≦ (2k +1) π/2n, θ and | Δ G |, are calculated by
θ=tan-1(ΔV/ΔH)+π/2
|ΔG|=(|ΔH|+|ΔV|)/2
Wherein, DeltaVAnd ΔHObtaining a vertical difference value and a horizontal difference value through a 3 multiplied by 3 Prewitt operator;
step 2.4: the characteristics are as follows: the Linearity (LIN) is calculated by,
wherein, PDd(i, j) denotes the value of the element of the n × n local direction co-occurrence matrix (DCOM) at a point (i, j) which is a distance d and has a direction code i and a direction code j along the boundary, we calculate H using differentiation and thresholdingDAnd PDdThe local direction co-occurrence matrix is defined as the relative frequency of two adjacent units, and is separated by a convenient distance d along the image, wherein the value is +1 when the local direction co-occurrence matrix is the same as the original direction, and the value is-1 when the local direction co-occurrence matrix is vertical to the original direction;
step 2.5: the characteristics are as follows: the Regularity (REG) is calculated by,
REG=1-r(σCON+σCOR+σDIR+σLIN)
wherein σXXXR is a normalization factor corresponding to the standard deviation of feature XXX;
step 2.6: the characteristics are as follows: the rough degree (ROU) is calculated by,
ROU=COA+CON
wherein COA and CON represent roughness and contrast, respectively;
the first type of characteristic Tamura characteristic is denoted as FTamuraI.e. FTamura={CON,COR,DIR,LIN,REG,ROU}。
The calculation of the feature of the central symmetric local binary pattern (CS-LBP) histogram of the critical region in step 2 comprises the following steps:
step 2.7: the centrosymmetric local binary pattern is calculated in such a way,
wherein R represents the radius, N represents the number of pixels captured in the neighborhood, T represents the threshold, CSLBPR,N,T(x, y) represents the value of the centrosymmetric local binary pattern feature with radius of R and neighborhood point number of N at the position (x, y);
step 2.8: extracting histogram features from the obtained centrosymmetric local binary pattern map;
let feature of the second kind be denoted as FCSLBPI.e., (28) extracted histogram features.
The gray histogram feature of the key area in the step 2 comprises the following steps:
step 2.9: converting the key area image into a gray image, and performing a gray value cutting process, namely, if the key area image is larger than a threshold value ThAll gray values of (1) are assigned as Ttru,
Wherein, Inorm(x, y) represents the grayscale value of the normalized key region at position (x, y);
step 2.10: calculating the gray level histogram of the key region image after the truncation processing, wherein the gray level distribution is [0, Ttru]Wherein T istruGenerally taking 180;
let feature type three be denoted as FHistI.e., (29) extracted histogram features.
In step 3, a STAR model is used for modeling the time sequence of the static features, and the method comprises the following steps:
step 3.1: the STAR model can be expressed as
Wherein, ai(i ═ 1,2,. cndot., P) and bi(i 1, 2.. P.) are all parameters to be estimated, ytIs the time series to be investigated, a0And b0Is a constant, P denotes the order of the investigated sequence, εtIs independent and equally distributed random variable, G (-) is a continuous smooth function which is divided into a logistic function and an exponential function,
G(yt-d;γ,c)=[1+exp(-γ×(yt-d-c))]-1;γ>0
G(yt-d;γ,c)=1-exp(-γ×(yt-d-c)2);γ>0
wherein, yt-dIs a threshold variable, d is a delay parameter, c is a threshold value between the two mechanisms;
step 3.2: determining a parameter P according to experience, and performing nonlinear least square estimation on the model to obtain an estimated value of the model parameter;
step 3.3: respectively modeling the time sequences of the three types of static features by using a STAR model, and taking the solution of the model as a final feature vector for depicting the dynamic features of the key region
The step 4 of judging whether the current video segment has the black smoke car or not comprises the following steps:
step 4.1: obtaining different model parameters by using different training samples, training an SVM model by using the parameters, and obtaining three SVM classifiers by using three types of feature training for classifying new estimation parameters;
step 4.2: respectively using the three models for classification of new features, and performing weighted fusion on classification results to obtain an identification result of the current key region;
wherein S iskRepresenting the classification result of the kth feature classifier;
step 4.3: and judging whether the current video segment has the black smoke car or not by analyzing the identification results of the continuous key areas.
Fig. 2 shows an example of a black smoke vehicle detected from a vehicle surveillance video using the present invention, white thin-line rectangular boxes representing bounding boxes of detected moving objects, gray rectangular boxes representing key regions, and the numerical values at the bottom of the gray rectangular boxes representing the probability that the region is a black smoke region.
It should be noted that the above-mentioned embodiments are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention, and all equivalent substitutions or substitutions made on the basis of the above-mentioned technical solutions belong to the scope of the present invention.
Claims (3)
1. A black smoke vehicle intelligent monitoring method based on a codebook and a smooth conversion autoregressive model is characterized by comprising the following steps:
(1) detecting a moving target in a video by using a Codebook model, and determining a key area behind the target;
(2) extracting Tamura characteristics, central symmetry local binary pattern CS-LBP histogram characteristics and gray level histogram characteristics of the key area, and combining to form three types of static characteristics of the key area;
(3) respectively modeling the time sequences of the three types of static features by utilizing a smooth transition autoregressive STAR model, and taking the solution of the model as a final feature vector for depicting the dynamic features of a key region;
(4) three SVM classifiers are obtained through three types of feature training, the three classification results are subjected to weighted fusion to obtain the recognition result of the current key area, and whether the current video segment has black smoke cars or not is judged through the analysis of the recognition results of a plurality of continuous key areas;
the Tamura characteristics of the key region in the step (2) comprise the following steps:
(21) the method is characterized in that: the roughness COA is calculated by the following method,
where m and n denote the width and height of the image, { E1,E2,...,Ek,...,ELDenotes the horizontal average intensity difference E at position (i, j) when k takes values in the set {1,2,3,4} throughk,h(i, j) and the difference E in vertical average intensityk,v(ii) a set of values of (i, j),
Ek,v(i,j)=|Ak(i,j+2k-1)-Ak(i,j-2k-1)|;
Ek,h(i,j)=|Ak(i+2k-1,j)-Ak(i-2k-1,j)|;
wherein F (i, j) represents the grayscale value at position (i, j), Ak(x1,x2) Indicates that the active window is 2k×2kK is the average value of pixels of 1,2,3,4, 5;
(22) the second characteristic: the contrast CON is calculated by, for example,
wherein σ represents the standard deviation, μ4Fourth moment of mean, parameter n0Set to 0.25;
(23) the characteristics are three: the direction DIR is calculated by a method comprising,
where φ represents the quantization direction code, p represents the histogram peak, and for any one peak p, WpRepresenting the quantization range contained by the peak, npIndicates the number of peaks in the histogram, phipRepresents WpQuantized values in the medium-maximum histogram, HDRepresenting a histogram, by calculating the number of points at which the gradient magnitude | Δ G | is greater than the threshold tho and the angle θ is within a certain range, HD(k) The number of points representing the point amplitude | Δ G | ≧ tho and the angle θ satisfying (2k-1) π/2n ≦ θ ≦ (2k +1) π/2n, θ and | Δ G |, are calculated by
θ=tan-1(ΔV/ΔH)+π/2;
|ΔG|=(|ΔH|+|ΔV|)/2;
Wherein, DeltaVAnd ΔHObtaining a vertical difference value and a horizontal difference value through a 3 multiplied by 3 Prewitt operator;
(24) the characteristics are as follows: the linearity LIN is calculated by the following method,
wherein, PDd(i, j) represents the element value of the n × n local direction co-occurrence matrix DCOM at a point (i, j) which is a distance d and has a direction code i and a direction code j along the boundary, and H is calculated by differentiation and thresholdingDAnd PDdThe local direction co-occurrence matrix is defined as the relative frequency of two adjacent units, which are separated by a convenient distance d along the image, and the value is +1 when the local direction co-occurrence matrix is the same as the original direction and-1 when the local direction co-occurrence matrix is perpendicular to the original direction;
(25) the characteristics are as follows: the regularity REG is calculated by the following method,
REG=1-r(σCON+σCOR+σDIR+σLIN);
wherein σXXXR is a normalization factor corresponding to the standard deviation of feature XXX;
(26) the characteristics are as follows: the rough ROU is calculated by the following method,
ROU=COA+CON;
wherein COA and CON represent roughness and contrast, respectively;
the first type of characteristic Tamura characteristic is denoted as FTamura,
I.e. FTamura={CON,COR,DIR,LIN,REG,ROU};
The calculation of the feature of the CS-LBP histogram of the central symmetry local binary pattern of the key region in the step (2) further comprises the following steps:
(27) the centrosymmetric local binary pattern is calculated in such a way,
wherein R represents the radius, N represents the number of pixels captured in the neighborhood, T represents the threshold, CSLBPR,N,T(x, y) represents the numerical value of the centrosymmetric local binary pattern feature with the radius of R and the number of neighborhood points of N at the position (x, y);
(28) extracting histogram features from the obtained centrosymmetric local binary pattern map;
let feature of the second kind be denoted as FCSLBPI.e., (28) extracted histogram features;
the gray histogram feature of the key area in the step (2) further comprises the following steps:
(29) converting the key area image into a gray image, and performing a gray value cutting process, namely, if the key area image is larger than a threshold value ThAll gray values of (1) are assigned as Ttru,
Wherein, Inorm(x, y) tableIndicating the gray value of the normalized key area at the position (x, y);
(210) calculating the gray level histogram of the key region image after the truncation processing, wherein the gray level distribution is [0, Ttru]Wherein T istruTaking 180;
let feature type three be denoted as FHistI.e., (29) extracted histogram features;
in the step (3), the STAR model is used for modeling the time series of the static features, and the method comprises the following steps:
(31) the STAR model is expressed as
Wherein, ai1,2, 1, P and biI 1,2, P being the parameter to be estimated, ytIs the time series to be investigated, a0And b0Is a constant, P denotes the order of the investigated sequence, εtIs independent and equally distributed random variable, G (-) is a continuous smooth function which is divided into a logistic function and an exponential function,
G(yt-d;γ,c)=[1+exp(-γ×(yt-d-c))]-1;γ>0
G(yt-d;γ,c)=1-exp(-γ×(yt-d-c)2);γ>0
wherein, yt-dIs a threshold variable, d is a delay parameter, c is a threshold value between the two mechanisms;
(32) determining a parameter P according to experience, and performing nonlinear least square estimation on the model to obtain an estimated value of the model parameter;
(33) and respectively modeling the time sequences of the three types of static features by using a STAR model, and taking the solution of the model as a final feature vector for depicting the dynamic features of the key region.
2. The intelligent black smoke vehicle monitoring method based on Codebook and smooth transition autoregressive model as claimed in claim 1, wherein the detection of the moving object in the video by using Codebook model in step (1) comprises the following points:
(11) the initialization of the codebook model and the characterization of the model,
c is1,c2,...cLFor a one-pixel codebook, the codebook for each pixel is set to null, i.e., L is 0, where L represents the number of codewords in each codebook, and each codeword c is set to zeroi1,2, L is a grey value viAnd hexahydric group auxiThe two parts are formed into a whole body,
wherein the content of the first and second substances,andrespectively representing the minimum and maximum grey values, f, in the code wordiIndicating the frequency of occurrence of the code word, λiIndicating the maximum time interval, p, during which the codeword does not occur during trainingiAnd q isiRespectively representing the time of the first and last occurrence of the codeword;
(12) the modeling of the codebook and the updating of the model,
i) let X be { X ═ X1,x2,...,xNIs a sequence of sample values for a pixel, xtIs the gray value at the moment t of the current pixel, for each x in the sampled value of each pixeltFinding out the code word c matched with it by using the following conditionm,
Wherein [ I ]low,Ihigh]Representing the range of variation of the grey value of each code word, epsilon1Is a threshold value, vmRepresenting a code wordcmCorresponding gray values;
ii) if the codebook is empty or no matching codeword is found, a new codeword is created,
if there is a code word cmThe code word is updated to be the code word when the two conditions are met
iii) after the training is over, calculating the maximum time interval during which each code word of the pixel does not reappear, and for each code word cii=1,2,...,L,λi=max{λi,(N-qi+pi-1) using λ to eliminate redundant code words, and obtaining an initial codebook M ═ c that best represents the real backgroundk|ck∈C,λk≤Tm},TmUsually half the number of training frames;
(13) the foreground object is extracted by the method of extracting,
for a new input pixel xtIf a matching codeword c satisfying the following condition can be found in the codebook MmThen xtForeground moving object pixels, otherwise background pixels,
wherein [ I ]low,Ihigh]Representing the range of variation of the grey value of each code word, epsilon2Is a threshold value, vmRepresents a code word cmTo the corresponding gray value.
3. The intelligent black smoke vehicle monitoring method based on the codebook and the smooth transition autoregressive model as claimed in claim 1, wherein the step (4) of determining whether there is a black smoke vehicle in the current video segment comprises the following steps:
(41) obtaining different model parameters by using different training samples, training an SVM model by using the parameters, and obtaining three SVM classifiers by using three types of feature training for classifying new estimation parameters;
(42) respectively using the three models for classification of new features, and performing weighted fusion on classification results to obtain an identification result of the current key region;
wherein S iskRepresenting the classification result of the kth feature classifier;
(43) and judging whether the current video segment has the black smoke car or not by analyzing the identification results of the continuous key areas.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810786681.7A CN109241824B (en) | 2018-07-17 | 2018-07-17 | Intelligent black smoke vehicle monitoring method based on codebook and smooth conversion autoregressive model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810786681.7A CN109241824B (en) | 2018-07-17 | 2018-07-17 | Intelligent black smoke vehicle monitoring method based on codebook and smooth conversion autoregressive model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109241824A CN109241824A (en) | 2019-01-18 |
CN109241824B true CN109241824B (en) | 2021-12-17 |
Family
ID=65072073
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810786681.7A Active CN109241824B (en) | 2018-07-17 | 2018-07-17 | Intelligent black smoke vehicle monitoring method based on codebook and smooth conversion autoregressive model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109241824B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110263867A (en) * | 2019-06-24 | 2019-09-20 | 株洲时代电子技术有限公司 | A kind of rail defects and failures classification method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103020980A (en) * | 2011-09-20 | 2013-04-03 | 佳都新太科技股份有限公司 | Moving target detection method based on improved double-layer code book model |
CN103310460A (en) * | 2013-06-24 | 2013-09-18 | 安科智慧城市技术(中国)有限公司 | Image characteristic extraction method and system |
CN103729862A (en) * | 2014-01-26 | 2014-04-16 | 重庆邮电大学 | Self-adaptive threshold value moving object detection method based on codebook background model |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2011223287A (en) * | 2010-04-09 | 2011-11-04 | Sony Corp | Information processor, information processing method, and program |
-
2018
- 2018-07-17 CN CN201810786681.7A patent/CN109241824B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103020980A (en) * | 2011-09-20 | 2013-04-03 | 佳都新太科技股份有限公司 | Moving target detection method based on improved double-layer code book model |
CN103310460A (en) * | 2013-06-24 | 2013-09-18 | 安科智慧城市技术(中国)有限公司 | Image characteristic extraction method and system |
CN103729862A (en) * | 2014-01-26 | 2014-04-16 | 重庆邮电大学 | Self-adaptive threshold value moving object detection method based on codebook background model |
Also Published As
Publication number | Publication date |
---|---|
CN109241824A (en) | 2019-01-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109753929B (en) | High-speed rail insulator inspection image recognition method | |
CN108304813B (en) | Method for intelligently identifying state of round traffic signal lamp | |
CN113421269A (en) | Real-time semantic segmentation method based on double-branch deep convolutional neural network | |
CN110110675B (en) | Wavelet domain fractal infrared cirrus cloud detection method fusing edge information | |
CN107679495B (en) | Detection method for movable engineering vehicles around power transmission line | |
CN112270247A (en) | Key frame extraction method based on inter-frame difference and color histogram difference | |
Zhang et al. | Vehicle recognition algorithm based on Haar-like features and improved Adaboost classifier | |
CN108416316B (en) | Detection method and system for black smoke vehicle | |
CN109190455B (en) | Black smoke vehicle identification method based on Gaussian mixture and autoregressive moving average model | |
CN109191495B (en) | Black smoke vehicle detection method based on self-organizing background difference model and multi-feature fusion | |
CN109460722B (en) | Intelligent license plate recognition method | |
CN114299002A (en) | Intelligent detection system and method for abnormal road surface throwing behavior | |
CN113177467A (en) | Flame identification method, system, device and medium | |
CN110060221B (en) | Bridge vehicle detection method based on unmanned aerial vehicle aerial image | |
CN106845458B (en) | Rapid traffic sign detection method based on nuclear overrun learning machine | |
CN111915583A (en) | Vehicle and pedestrian detection method based on vehicle-mounted thermal infrared imager in complex scene | |
Du et al. | Research on an efficient method of license plate location | |
Achmad | Gaussian Mixture Models optimization for counting the numbers of vehicle by adjusting the Region of Interest under heavy traffic condition | |
Ogunrinde et al. | A review of the impacts of defogging on deep learning-based object detectors in self-driving cars | |
CN109241824B (en) | Intelligent black smoke vehicle monitoring method based on codebook and smooth conversion autoregressive model | |
CN111242891A (en) | Method for identifying and classifying surface defects of steel rails | |
CN110751667A (en) | Method for detecting infrared dim small target under complex background based on human visual system | |
CN113705433A (en) | Power line detection method based on visible light aerial image | |
Kaliyev et al. | Computer Vision-based Fire Detection using Enhanced Chromatic Segmentation and Optical Flow Model | |
CN107341456B (en) | Weather sunny and cloudy classification method based on single outdoor color image |
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 |