CN104166841B - The quick detection recognition methods of pedestrian or vehicle is specified in a kind of video surveillance network - Google Patents
The quick detection recognition methods of pedestrian or vehicle is specified in a kind of video surveillance network Download PDFInfo
- Publication number
- CN104166841B CN104166841B CN201410356465.0A CN201410356465A CN104166841B CN 104166841 B CN104166841 B CN 104166841B CN 201410356465 A CN201410356465 A CN 201410356465A CN 104166841 B CN104166841 B CN 104166841B
- Authority
- CN
- China
- Prior art keywords
- image
- pixel
- edge
- lbp
- template
- 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
Abstract
A kind of quick detection recognition methods the present invention relates to specify pedestrian or vehicle in video surveillance network.According to the To Template that user provides, to specifying target that accurately and quickly to be detected identification in Internet video is monitored.The present invention is normalized various composite characters of change of scale and calculation template image to To Template image first；Then background modeling is carried out to each monitor video using mixed Gauss model, parallel processing extracts sport foreground using area filtering and morphology post processing；Change of scale is normalized to all sport foregrounds for meeting preliminary screening condition and various composite characters are calculated, the weighting similarity distance between each sport foreground and To Template is calculated afterwards, to be returned less than decision threshold and with the top n object of To Template image weighting similarity distance minimum, as detection recognition result.The present invention substantially increases the processing speed of algorithm by methods such as parallel processing and frame-skippings on the basis of detection recognition accuracy is ensured.
Description
Technical field
The present invention relates to technical field of video image processing, specified pedestrian or car in specially a kind of video surveillance network
Quick detection recognition methods.
Background technology
In field of intelligent monitoring, to monitor video network in specified pedestrian or vehicle detected and recognized and can be helped
Public security organ quickly determines time of occurrence and the position of suspect or suspected vehicles, accelerates case investigation efficiency.But due to reality
Existed in application scenarios illumination condition change, exist between different monitoring camera parameter differences, moving object (pedestrian,
Vehicle etc.) there is attitudes vibration and mutually block, and the size of pedestrian or vehicle in monitor video generally smaller, such as people
The detailed information such as face or car plate can not be guaranteed, thus how multi-cam monitoring network in quickly and accurately detect with
Identify the focus and difficult point of specified pedestrian or vehicle always computer vision field research.
Oreifej is published in 2010《IEEE Comptuer Society Conference on Computer
Vision and Pattern Recognition》(computer society of International Electrical IEEE computer vision with
Pattern-recognition meeting) technology " the Human identity recognition in that deliver on technology collection page 709 to 716
A kind of taking photo by plane across regarding based on vote by ballot is proposed in aerial images " (pedestrian's identification in Aerial Images) article
Pedestrian recognition method under frequency complex background, this technology is first with HoG (Histograms of Oriented Gradient
Gradient orientation histogram) pedestrian area in feature detection video, the pedestrian's set occurred in video sequence is constituted, referred to as
Candidate gathers, and candidate is thrown by characteristic similarity by pre-prepd target pedestrian's characteristic set for training
Ticket, gained vote highest pedestrian is most suspicious object.The algorithm preferably resolves that image resolution ratio is relatively low, pedestrian's attitude is changeable
Problem is recognized Deng the target pedestrian under complex environment.But the method in video image because multiple dimensioned being repeatedly scanned with will one by one examine
Pedestrian area is surveyed, and needs multiple To Template images to gather as voter, therefore all deposited in speed and practical feasibility
In deficiency.
Because monitor video generally has fixed viewpoint so that the method for splitting sport foreground by background modeling can be with
Effectively implement, differentiate that such method is in fixed viewpoint scene compared to multiple dimensioned scanning is carried out to video image using HoG features
Under it is quick and efficient to the detection of moving object segmentation, but the method will no longer be applicable when camera has mobile.
Reduce training link be also target detection identification in difficult point, at present some be based on artificial neural network and support to
Although the detection recognizer recognition accuracy of amount machine is high, amount of calculation is larger, and needs to be provided previously by a large amount of To Template figures
It is not high in the less field application value of the To Template information such as criminal investigation as training.
The above method respectively has advantage and disadvantage, how the above method effectively to be combined for application-specific scene, current
Study and insufficient.This promotes to find a kind of rational method framework of criminal investigation work network suitable for monitoring, in ensuring method
While accuracy rate, method speed is improved, reach real-time processing.
The content of the invention
The present invention provides the quick detection recognition methods of specified pedestrian or vehicle in a kind of video surveillance network, existing to solve
There is technology for detection speed slower, and need the problem trained under the enterprising line of training set in advance mostly.
To reach above-mentioned purpose, the present invention is using the following technical scheme counted：Pedestrian is specified in a kind of video surveillance network
Or the quick detection recognition methods of vehicle, comprise the following steps：
Step 1：Taken from the picture comprising specified pedestrian or vehicle or video center by way of rectangle frame is demarcated first
Identification target to be detected；
Step 2：To Template image to selecting is normalized change of scale, and calculates the various of To Template image
Composite character；
Step 3：For the monitor video that each camera shoots opens up independent thread, to each camera in monitoring network simultaneously
Row treatment, background modeling is carried out using gauss hybrid models respectively to the monitoring scene residing for each camera；
Step 4：The moving foreground object in each camera is extracted using area filtering and morphology post processing；By setting
Determine the moving target area threshold moving target too small to filter area, pedestrian or vehicle are not belonging in scene to filter out
Other moving objects；
Step 5：To each monitor video frame-skipping extract sport foreground, in video surveillance network it is all meet preliminary treatment will
The moving object asked is normalized change of scale and calculates various composite characters, is weighted phase with To Template image afterwards
Like property measurement；
Step 6：According to the similitude size with template image, the detected motion in each monitor video in network will be monitored
The information of object is dynamically put into the result vector V of designated lengthresult, after all monitor videos are traveled through, return to the result vector
As final detection recognition result.
Further, the various composite characters described in step 2 are main color characteristic, block margin direction histogram spy
Levy, piecemeal HSV histogram features, HoG features and invariable rotary LBP features of equal value.It implements process：
Step 2.1：The generating process of wherein domain color feature is：
1. the To Template image after Normalized Scale is converted is transformed into HSV space from rgb space, only extracts image and exists
Chrominance component in HSV space；
2. the span of chrominance component is divided into 8 intervals, by the chrominance component of To Template image according to this 8
Interval projection turns into the value chist in i-th interval of 8 dimension H histogram of component chist, wherein chistiObtained by following formula：
Wherein hx,yRefer to the chrominance component value of image (x, y) coordinate position pixel, Rect refer to To Template image or
The foreground area that person's sport foreground is detected, δi(hx,y) be defined as follows：
3. the interval average tone value c of 8 tones is calculated respectivelyi, and 8 dimension histograms are normalized, obtainP is chosen from 8 dimension normalization histogramsiMaximum 3-dimensional, uses vector vdcPreserve the c corresponding to this 3-dimensionaliWith
piConstitute the domain color feature of the target.Therefore domain color feature vdcIt is made up of 6 dimensions altogether, accounting is maximum in having corresponded to target image
3 color components tone value and percentage vdc=[c1,p1,c2,p2,c3,p3]；
Step 2.2：The generating process of wherein block margin direction histogram is：
1. the target image after Normalized Scale is converted first is converted into gray level image by RGB color image, and splits
Into 4 × 4 totally 16 pieces；
2. gray level image is filtered using Sobel horizontal edges detective operators and Sobel vertical edge detective operators,
The horizontal edge intensity and vertical edge intensity of each pixel are obtained, and the edge direction of each pixel is obtained using them
ex,yAnd edge strength | ex,y|；
3. pixel is divided into five types according to edge direction and intensity：Horizontal edge, vertical edge, 45 ° of edges,
135 ° of edges and directionless edge；Edge orientation histogram is built using four kinds of edges above；
4. directionless edge threshold T is sete, in each piecemeal, T is more than to all edge strengthsePixel according to side
Edge direction carries out statistics with histogram, and edge orientation histogram is tieed up in generation 4；4 × 4 piecemeal symbiosis tie up edge direction Nogatas into 64
Figure, local edge direction feature ehist is saved as by this 64 dimensional vectorl；Again by 4 dimension edge orientation histograms in 16 piecemeals
Added up, constituted the global edge direction characteristic ehist of 4 dimensions of view picture target imageg；They are merged into 68 dimensional vectors, are protected
Save as block margin direction histogram feature ved；
Step 2.3：The wherein histogrammic generating process of piecemeal HSV is：
1. To Template image or sport foreground image after Normalized Scale is converted are projected to by rgb color space
HSV color spaces, and level is divided into 4 pieces from top to bottom by image；
2. the HSV histograms in 4 piecemeals are calculated respectively, wherein chrominance component is divided into 16 intervals, saturation degree component and
Luminance component is respectively divided into 4 intervals, and each piecemeal finally gives 24 dimension histogram features；4 24 dimension histograms are spliced to form
96 dimension piecemeal HSV histograms of entire image；
Step 2.4：The generating process of wherein HoG features is：
1. the To Template image after Normalized Scale is converted is converted into gray level image, and gray level image is carried out
Gamma is corrected, and reduces the influence of local shades and illumination variation to characteristic extraction procedure；
2. convolution algorithm is done to image with [- 1,0,1] gradient operator, obtains the horizontal gradient G of each pixelx(x, y),
Use [1,0, -1] againTGradient operator does convolution algorithm to image, obtains the vertical gradient G of each pixely(x,y)；And utilize water
Flat ladder degree and vertical gradient calculate gradient magnitude and gradient direction：
3. the cell factory of multiple 8 × 8 pixels is divided the image into, gradient direction is averagely divided into 9 by span
Individual interval, according to the gradient direction of each pixel in cell factory, histogram of gradients is tieed up in generation 9；
4. setting every 4 adjacent cell factory composition one piecemeal, i.e. each piecemeal has 16 × 16 pixels, and between piecemeal
Can be overlapped；By 49 dimension histogram of gradients series connection of cell factory in each piecemeal, the piecemeal description vectors of 36 dimensions are obtained；
5. the window of setting 64 × 64, window is vertically slided from top to bottom, is slided at intervals of 8 pixels,
The description vectors of all piecemeals that window is included are connected, and constitute the description vectors of whole window, then by all windows
Description vectors connected, finally give HoG characteristic vectors vhog；
Step 2.5：The generating process of invariable rotary LBP features wherein of equal value is：
1. To Template image or sport foreground image after Normalized Scale is converted are converted into gray level image, for figure
In each pixel, compare itself and 8 gray value sizes of neighbor pixel around, neighbours' point gray value is more than the pixel then
1 is set to, otherwise is then set to 0；8 numeric strings are unified into 82 systems in the direction of the clock since 12 o'clock position again
Number, computing formula is as follows：
Wherein, P is neighbours' number, and R is radius size, gyAnd gcThe respectively gray value and center picture of neighbor pixel point
The gray value of vegetarian refreshments；
2. each the 82 system numbers that will be obtained are joined end to end, and form an annular, and the circular clockwise is rotated 7 times,
8 groups 82 system numbers can be obtained altogether, 8 group of 22 minimum system number of system number intermediate value is chosen, and the number is pixel correspondence
Invariable rotary LBP values；
3. the invariable rotary LBP values that will 2. be obtained by step are divided into two classes, and wherein 0-1 conversion times are not more than 2 times and return
It is a class, referred to as consistent LBP operators；Remaining 2 system number is all classified as another kind of；By above-mentioned classification, consistent invariable rotary
LBP operators have 9 kinds, and non-uniform LBP operators have a kind；
4. the fritter of multiple 16 × 16 pixels is divided the image into, pixel LBP values in each fritter are calculated, one is constituted
The invariable rotary LBP histograms of equal value of 10 dimensions；The LBP histograms of all fritters are connected, the equivalence of 840 dimensions is finally given
Invariable rotary LBP features vLBP；
Further, before the filtering of utilization area and morphology post processing described in step 4 are extracted in each camera
Scape target, it is implemented including following sub-step：
Step 4.1：Area filtering is carried out to the binary image after mixed Gaussian background modeling and thresholding treatment；
Step 4.2：3 expansive workings are carried out with 3 × 3 templates to binary image；
Step 4.3：1 etching operation is carried out with 3 × 3 templates again to binary image；
Step 4.4：Finally carry out an area filtering again to binary image.
Area filtering described in step 4 refers to the fritter that binary image is divided into multiple 4 × 4 pixels, if certain
When the number of foreground pixel point is less than or equal to 3 in individual fritter, then all pixels point in the fritter is defined as background pixel point, instead
The foreground pixel point then retained in the fritter；
The moving target too small to filter area by setting moving target area threshold described in step 4, it is specific
Implementation process is：All of connected region in binary image is found out after treatment, each connected region foreground pixel point is calculated
Number, setting area threshold ThcIt is the 1/400 of video frame images area, if the area of connected region is more than ThcThen retain connection
Region, and the minimum rectangle frame comprising the connected region is returned as sport foreground region；If connected region area is less than or equal to
ThcThe connected region is then set to background area.
Further, sport foreground is extracted to each monitor video frame-skipping described in step 5, it implements process and is：
According to user set frame-skipping number F, every F frames again will in step 4 extract sport foreground object one by one with To Template image
It is weighted similarity measurement；
Weighting similarity measurement described in step 5, it implements process and is：First respectively calculate sport foreground object with
The domain color characteristic distance D of To Template imagedc, block margin direction histogram characteristic distance Ded, piecemeal HSV histograms it is special
Levy apart from Dhsv, HoG characteristic distances DhogWith invariable rotary LBP characteristic distances D of equal valueLBP, then the distance of five kinds of features is returned
One changes weighting similarity measurement, obtains weighting similarity distance Dall：
Wherein α, β, γ, λ,It is normalized weighing factors, for balancing influence of each feature to overall similarity measurement,
Tested to adjust in advance according to specific environment.
Further, the information that will monitor the detected moving object in network in each monitor video described in step 6 is moved
State is put into the result vector V of designated lengthresult, it is implemented including following sub-step：
Step 6.1：Create a vectorial V with designated length Nresult, wherein N is the monitor video net for needing to return
With specified pedestrian or the immediate object number of car modal image in network；
Step 6.2：When there is new sport foreground to be detected in monitor video network, V is first determined whetherresultMiddle preservation
Object number；
Step 6.3：If VresultThe object number of middle preservation is less than N, when object and To Template image weighting similitude away from
From DallMeet similarity threshold Th set in advanceallWhen, then the residing camera of the object is numbered, frame number occur, occur
The information such as the subject image after rectangle frame coordinate position in the frame and Normalized Scale conversion are retained in vectorial Vresult
In；
Step 6.4：If VresultThe object number of middle preservation has reached N, and when object and the weighting phase of To Template image
Like property apart from DallMeet similarity threshold Th set in advanceallWhen, then the weighting similarity distance D for being obtained with the objectallWith
VresultIn the weighting similarity distance that obtains of the object maximum with To Template distanceCompare, if
Then V is substituted with the objectresultIn the object maximum with To Template distance, and by VresultN number of object of middle preservation according to
To Template similarity distance is resequenced from small to large；
Step 6.5：After the monitor video that all cameras shoot in traversal monitoring network, the V of returnresultProtected in vector
Deposit in whole monitoring network with specified pedestrian or the immediate N number of object of car modal image, and comprising residing for each of which
Camera numbering, occur frame number, there is the thing after rectangle frame coordinate position in the frame and Normalized Scale conversion
Body image information.
Principle of the invention is that the pedestrian or vehicle in monitor video network must be in certain time periods in the presence of the mistake moved
Cheng Caineng enters and leaves monitoring range, and monitoring camera all has fixed visual angle, so if will to specify pedestrian or
Vehicle is used for quickly detecting identification, only need to respectively carry out background modeling to each camera in network, in extracting each camera
The foreground object of motion, by after preliminary screening, thread individually being opened up to each camera, with specified pedestrian or the mould of vehicle
Plate image carries out multiple features Weighted Fusion similarity measurement with the moving object after primary election, then to the similitude in each camera
Result carries out collecting sequence, you can entirely monitored most like with specified pedestrian or vehicle top n moving object in network,
As the result that quick detection is recognized；Due to the sport foreground included in adjacent several frames in monitor video do not have typically it is too big
Change, therefore the present invention can carry out frame-skipping treatment during background modeling and similarity measurement, be carried out once every F frames
Sport foreground similarity measurement, the arithmetic speed of system is further speeded up with this.
The present invention has the advantages that：
1) compared with prior art, fixed present invention utilizes monitoring camera visual angle and there will necessarily be fortune with pedestrian or vehicle
Dynamic this two dot characteristics of process, target detection scope is primarily determined that using the method for Analysis on Prospect, compared to multiple dimensioned searching algorithm,
The target zone of detection identification can be effectively reduced, the efficiency of detection identification is improved；
2) present invention comprehensive utilization various features, at the same consider to specify pedestrian or vehicle and each sport foreground color,
Many degree of similarity such as shape, texture, and fusion is weighted to multi-aspect information；
3) present invention only needs to simple target template as input, it is not necessary to provide the database use for being detected target
In training, the method feasibility in actual applications is considerably increased；
4) invention introduces multi-cam parallel processing and frame-skipping treatment scheduling algorithm, so that the present invention is in complex scene
In both accurately can robustly complete to specify the detection identification of pedestrian or vehicle, the operation time of method can be effectively reduced again.
Brief description of the drawings
Fig. 1 is overall flow schematic diagram of the invention；
Fig. 2 (a) is monitoring scene schematic diagram of the invention；
Fig. 2 (b) is mixed Gaussian background modeling result figure of the invention；
Fig. 2 (c) is that morphology of the invention post-processes result figure；
Fig. 2 (d) is motion foreground segmentation final result figure of the invention；
Fig. 3 is the specified To Template image schematic diagram of the present invention；
Fig. 4 is the specified target quick detection recognition result figure of the present invention；
Specific embodiment
With reference to specific embodiment, this invention is expanded on further.The present embodiment is premised on technical solution of the present invention
Under implemented, give detailed implementation method and specific operating process, but protection scope of the present invention be not limited to it is following
Embodiment.
Embodiment
The selected software environment of the present invention：Windows7.0 operating systems, VS2010 development platforms, C/C++ exploitation languages
Speech, OpenCV2.3.1 kits.
The present embodiment is processed certain monitoring network.Video scene is complex in the monitoring network, there are a large amount of fortune
Had between animal body, and moving object and mutually blocked, also due to parameter and position relationship, have illumination between each camera
Change and aberration.
See Fig. 1, the technical solution adopted in the present invention is：Specified pedestrian's or vehicle in a kind of video surveillance network
Quick detection recognition methods, comprises the following steps：
Step 1：Taken from the picture comprising specified pedestrian or vehicle or video center by way of rectangle frame is demarcated first
Identification target to be detected；
User is provided comprising the picture or monitor video data for specifying pedestrian or vehicle, using rectangle collimation mark by man-machine interface
Fixed method, the location of pedestrian or vehicle are determined from the picture or video frame images center fetching, and rectangle frame is included
Region is individually divided into piece image.
Step 2：To Template image to selecting is normalized change of scale, and calculates the various of To Template image
Composite character；
It is 64 × 128 pixel sizes by To Template image normalization change of scale, To Template image is calculated respectively
Domain color feature, block margin direction histogram feature, piecemeal HSV histogram features, HoG features and invariable rotary LBP of equal value
Feature, with many-sided to specifying pedestrian or vehicle to express from color, shape, texture etc..
Step 2.1：The generating process of wherein domain color feature is：
1. the To Template image after Normalized Scale is converted is transformed into HSV space from rgb space, only extracts image and exists
Chrominance component in HSV space；Domain color feature is calculated equivalent to the color group for only considering target only with chrominance component
Into, so that feature is insensitive to the illumination variation and saturation difference of different cameras, with the robustness of this Enhanced feature.
2. the span of chrominance component is divided into 8 intervals, by the chrominance component of To Template image according to this 8
Interval projection turns into the value chist in i-th interval of 8 dimension H histogram of component chist, wherein chistiObtained by following formula：
Wherein hx,yRefer to the chrominance component value of image (x, y) coordinate position pixel, Rect refer to To Template image or
The foreground area that person's sport foreground is detected.δi(hx,y) be defined as follows：
3. the interval average tone value c of 8 tones is calculated respectivelyi, and 8 dimension histograms are normalized, obtainP is chosen from 8 dimension normalization histogramsiMaximum 3-dimensional, uses vector vdcPreserve the c corresponding to this 3-dimensionaliWith
piConstitute the domain color feature of the target.Therefore domain color feature vdcIt is made up of 6 dimensions altogether, accounting is maximum in having corresponded to target image
3 color components tone value and percentage vdc=[c1,p1,c2,p2,c3,p3]
Two domain color characteristic vectors are given, then the distance between they DdcIt is defined as follows：
A in above formulai,jFor characterizing color component ciAnd cjSimilarity degree, it passes through below equation and is defined：
T in this exampled=25, dmax=50.
Step 2.2：The generating process of wherein block margin direction histogram is：
1. the target image after Normalized Scale is converted first is converted into gray level image by RGB color image, and splits
Into 4 × 4 totally 16 pieces.
2. gray level image is filtered using Sobel horizontal edges detective operators and Sobel vertical edge detective operators,
The horizontal edge intensity and vertical edge intensity of each pixel are obtained, and the edge direction of each pixel is obtained using them
ex,yAnd edge strength | ex,y|。
3. pixel is divided into five types by conventional domain color feature according to edge direction and intensity：Horizontal edge, hang down
Straight edge, 45 ° of edges, 135 ° of edges and directionless edges；Wherein directionless edge refers to edge strength less than a certain specific
The pixel of threshold value；Because directionless edge pixel point generally accounts for significant proportion in entire image, therefore in the present invention only
Consider four kinds of edges above to constitute edge orientation histogram.
4. directionless edge threshold T is sete, T in this examplee=100.It is big to all edge strengths in each piecemeal
In TePixel carry out statistics with histogram according to edge direction, edge orientation histograms are tieed up in generation 4.4 × 4 piecemeal symbiosis into
64 dimension edge orientation histograms, local edge direction feature ehist is saved as by this 64 dimensional vectorl；Again by 4 in 16 piecemeals
Dimension edge orientation histogram is added up, and constitutes the global edge direction characteristic ehist of 4 dimensions of view picture target imageg；They are closed
And into 68 dimensional vectors, save as block margin direction histogram feature ved。
Two block margin direction histogram features are given, then the distance between they DedIt is defined as：
Compared to local component, the present invention imparts bigger weight to global component.
Step 2.3：The wherein histogrammic generating process of piecemeal HSV is：
1. To Template image or sport foreground image after Normalized Scale is converted are projected to by rgb color space
HSV color spaces, and level is divided into 4 pieces from top to bottom by image.
2. the HSV histograms in 4 piecemeals are calculated respectively, wherein chrominance component is divided into 16 intervals, saturation degree component and
Luminance component is respectively divided into 4 intervals, therefore each piecemeal finally gives 24 dimension histogram features.By 4 24 dimension histogram splicings
Constitute 96 dimension piecemeal HSV histograms of entire image.
In order to adapt to the application scenarios of vehicle and pedestrian detection identification, i.e., in To Template or sport foreground image topmost
May include some background elements with bottom, and center section is mainly the important information such as vehicle body or pedestrian's clothes, this reality
Example measure two piecemeal HSV histogram features between apart from when impart bigger weight to middle two-layer piecemeal.
Two piecemeal HSV histogram features are given, apart from D between themhsvIt is defined as follows：
Dhsv(vhsv,v′hsv)=0.8d (hist1,hist1′)+1.2·d(hist2,hist′2)+1.0·d(hist3,
hist3′)+0.8·d(hist4,hist′4)
Wherein d (hist, hist ') is defined as card side's distance in this example：
Step 2.4：The generating process of wherein HoG features is：
1. the To Template image after Normalized Scale is converted is converted into gray level image, and gray level image is carried out
Gamma is corrected, and reduces the influence of local shades and illumination variation to characteristic extraction procedure.
2. convolution algorithm is done to image with [- 1,0,1] gradient operator, obtains the horizontal gradient G of each pixelx(x, y),
Use [1,0, -1] againTGradient operator does convolution algorithm to image, obtains the vertical gradient G of each pixely(x,y)；And utilize water
Flat ladder degree and vertical gradient calculate gradient magnitude and gradient direction：
3. the cell factory of multiple 8 × 8 pixels is divided the image into, each image will be divided into 128 in this example
Individual cell factory, gradient direction is averagely divided into 9 intervals by span, according to the ladder of each pixel in cell factory
Histogram of gradients is tieed up in degree direction, generation 9.
4. setting every 4 adjacent cell factory composition one piecemeal, i.e. each piecemeal has 16 × 16 pixels, and between piecemeal
Can be overlapped；By 49 dimension histogram of gradients series connection of cell factory in each piecemeal, the piecemeal description vectors of 36 dimensions are obtained.
5. the window of setting 64 × 64, window is vertically slided from top to bottom, is slided at intervals of 8 pixels,
The description vectors of all piecemeals that window is included are connected, and constitute the description vectors of whole window, then by all windows
Description vectors connected, finally give HoG characteristic vectors vhog；The HoG features generated in the present embodiment have 15876 dimensions,
The distance between Hog characteristic vectors measurement uses card side's distance：
Step 2.5：The generating process of invariable rotary LBP features wherein of equal value is：
1. To Template image or sport foreground image after Normalized Scale is converted are converted into gray level image, for figure
In each pixel, compare itself and 8 gray value sizes of neighbor pixel around, neighbours' point gray value is more than the pixel then
1 is set to, otherwise is then set to 0, then 8 numeric strings are unified into 82 systems in the direction of the clock since 12 o'clock position
Number.Computing formula is as follows：
Wherein, P is neighbours' number, and it is to be set to 1, g in radius size, this example that 8, R is set in this exampleyAnd gcIt is respectively adjacent
Occupy the gray value of pixel and the gray value of center pixel.
2. each the 82 system numbers that will be obtained are joined end to end, and form an annular, and the circular clockwise is rotated 7 times,
8 groups 82 system numbers can be obtained altogether, 8 group of 22 minimum system number of system number intermediate value is chosen, and the number is pixel correspondence
Invariable rotary LBP values；By the step, possible 2 system output is by 2836 kinds are reduced to, and the LBP values for obtaining are to image
Rotation there is robustness.
3. the invariable rotary LBP values that will 2. be obtained by step are divided into two classes, and wherein 0-1 conversion times are not more than 2 times and return
It is a class, referred to as consistent LBP operators；Remaining 2 system number is all classified as another kind of；By above-mentioned classification, consistent invariable rotary
LBP operators have 9 kinds, and non-uniform LBP operators have a kind.
4. the fritter of multiple 16 × 16 pixels is divided the image into, pixel LBP values in each fritter are calculated, one is constituted
The invariable rotary LBP histograms of equal value of 10 dimensions；The LBP histograms of all fritters are connected, the equivalence of 840 dimensions is finally given
Invariable rotary LBP features vLBP；The distance between invariable rotary LBP features of equal value DLBPIt is same to be measured using card side's distance：
Step 3：For the monitor video that each camera shoots opens up independent thread, to each camera in monitoring network simultaneously
Row treatment, background modeling is carried out using gauss hybrid models respectively to the monitoring scene residing for each camera；
Each camera in monitoring network has fixed video scene, the gray scale of each pixel in video scene
Value can be described with mixed Gauss model
Wherein, η is Gaussian probability-density function, and K is the number upper limit of Gaussian function in mixed Gauss model, in this example
Middle K is set as 5,WithRespectively k-th Gauss model is in the weight at t frame moment, average and variance.
The mixed Gaussian background modeling process of each camera is by the way that individually opening up thread realizes parallel computation in the present invention.
Comprising the following steps that for background modeling is carried out using gauss hybrid models：
1. the mixed Gauss model of each pixel is initialized with the gray value of each pixel of frame of monitor video first.Now
Mixed Gauss model only one of which Gaussian function is initialised, and its average is the gray value of current pixel, and variance is set to
Fixed value σ2=30, the weights of Gauss are set to 0.05.
2. when a new two field picture is read in, whether each Gaussian function is checked by Gaussian function weights order from large to small
Gray value with this pixel matches.I.e. grey scale pixel value is no more than Th with the difference of the Gaussian function averaged=2.5 σ=
13.69.If finding the Gaussian function of matching, step is jumped to 3..If this gray scale and any one Gaussian function are not
Match somebody with somebody, then a new Gaussian function is 1. reinitialized according to step.When the Gaussian function that there is no initializtion in mixed model
When, directly initialized with new Gaussian function；When K Gaussian function of setting is all used, then with this new Gaussian function
Count to replace the Gaussian function of weights minimum in current mixed model.
3., it is necessary to each has been used in mixed model after current pixel gray value corresponding Gaussian function is determined
The weights of Gaussian function, average and variance be updated.The modeling of background needs the regular hour to accumulate with renewal, it is stipulated that this
Time window length L=200.When the frame number that video reads in is less than 200, Gaussian function more new formula is：
Wherein, N is frame number, ωkFor recording sequence number of k-th Gaussian function in the arrangement of weights descending.It is two-valued function, works as ωkWhen Gaussian function sequence number with matching is identical, its value is 1, is otherwise 0.
After frame number over-time length of window L, more new formula is：
After renewal is finished, then the weights of each Gaussian function in mixed Gauss model are normalized.
4. each Gaussian function according to the descending sequence of weight after its renewal, determine that weight is added more than Thw=
0.7 preceding B Gaussian function is the Gaussian function for describing background.If the Gaussian function matched with current pixel come it is first B,
Judge that the pixel is background pixel, conversely, being then foreground pixel.
Step 4：The moving foreground object in each camera is extracted using area filtering and morphology post processing；By setting
Determine the moving target area threshold moving target too small to filter area, pedestrian or vehicle are not belonging in scene to filter out
Other moving objects；
The binary image obtained to mixed Gauss model carries out area filtering and Morphological scale-space can remove noise and
Hole, it is implemented including following sub-step：
Step 4.1：Area filtering is carried out to the binary image after mixed Gaussian background modeling and thresholding treatment；
Step 4.2：3 expansive workings are carried out with 3 × 3 templates to binary image；
Step 4.3：1 etching operation is carried out with 3 × 3 templates again to binary image；
Step 4.4：Finally carry out an area filtering again to binary image.
Wherein area filtering refers to the fritter that binary image is divided into multiple 4 × 4 pixels, if before in certain fritter
When the number of scene vegetarian refreshments is less than or equal to 3, then all pixels point in the fritter is defined as background pixel point, otherwise then retain and be somebody's turn to do
Foreground pixel point in fritter；
The moving target too small to filter area by setting moving target area threshold described in step 4, it is specific
Implementation process is：All of connected region in binary image is found out after treatment, each connected region foreground pixel point is calculated
Number, setting area threshold ThcIt is the 1/400 of video frame images area, if the area of connected region is more than ThcThen retain connection
Region, and the minimum rectangle frame comprising the connected region is returned as sport foreground region；If connected region area is less than or equal to
ThcThe connected region is then set to background area.
Step 5：To each monitor video frame-skipping extract sport foreground, in video surveillance network it is all meet preliminary treatment will
The moving object asked is normalized change of scale and calculates various composite characters, is weighted phase with To Template image afterwards
Like property measurement；
Because the video sequence each second in monitor video network includes 25 frames, thus no matter pedestrian or vehicle is adjacent
Can all repeat in several frames, in order to accelerate algorithm speed, present invention introduces frame skipping strategy, i.e., the frame-skipping number for being set according to user
F, similarity measurement is weighted every F frames with To Template image one by one by the sport foreground object extracted in step 4 again.This
F=3 is set in example.
Sport foreground image is normalized change of scale, 64 × 128 pixel images are converted into, and according in step 2
Method calculate respectively the domain color feature of sport foreground, block margin direction histogram feature, piecemeal HSV histogram features,
HoG features and invariable rotary LBP features of equal value.
For described weighting similarity measurement, it implements process and is：First it is utilized respectively distance degree between each feature
The definition of amount calculates the domain color characteristic distance D of sport foreground object and To Template imagedc, block margin direction histogram it is special
Levy apart from Ded, piecemeal HSV histogram features are apart from Dhsv, HoG characteristic distances DhogWith invariable rotary LBP characteristic distances D of equal valueLBP,
The distance of five kinds of features is normalized weighting similarity measurement again, weighting similarity distance D is obtainedall：
Wherein α, β, γ, λ,It is normalized weighing factors, for balancing influence of each feature to overall similarity measurement,
Tested to adjust in advance according to specific environment.Their value is respectively 10,1,5,1.2 and 1 in this example.
Step 6：According to the similitude size with template image, the detected motion in each monitor video in network will be monitored
The information of object is dynamically put into the result vector V of designated lengthresult, after all monitor videos are traveled through, return to the result vector
As final detection recognition result.
Further, the information that will monitor the detected moving object in network in each monitor video described in step 6 is moved
State is put into the result vector V of designated lengthresult, it is implemented including following sub-step：
Step 6.1：Create a vectorial V with designated length Nresult, wherein N is the monitor video net for needing to return
With specified pedestrian or the immediate object number of car modal image in network；
Step 6.2：When there is new sport foreground to be detected in monitor video network, V is first determined whetherresultMiddle preservation
Object number；
Step 6.3：If VresultThe object number of middle preservation is less than N, when object and To Template image weighting similitude away from
From DallMeet similarity threshold Th set in advanceallWhen, then the residing camera of the object is numbered, frame number occur, occur
The information such as the subject image after rectangle frame coordinate position in the frame and Normalized Scale conversion are retained in vectorial Vresult
In；
Step 6.4：If VresultThe object number of middle preservation has reached N, and when object and the weighting phase of To Template image
Like property apart from DallMeet similarity threshold Th set in advanceallWhen, then the weighting similarity distance D for being obtained with the objectallWith
VresultIn the weighting similarity distance that obtains of the object maximum with To Template distanceCompare, if
Then V is substituted with the objectresultIn the object maximum with To Template distance, and by VresultN number of object of middle preservation according to
To Template similarity distance is resequenced from small to large；
Step 6.5：After the monitor video that all cameras shoot in traversal monitoring network, the V of returnresultProtected in vector
Deposit in whole monitoring network with specified pedestrian or the immediate N number of object of car modal image, and comprising residing for each of which
Camera numbering, occur frame number, there is the thing after rectangle frame coordinate position in the frame and Normalized Scale conversion
Body image information.
N=4, and Th in this exampleall=50.
Implementation result
According to above-mentioned steps, to the video in monitoring network specify quick detection and the identification of pedestrian or vehicle.Fig. 2
The result figure in each stage of motion foreground segmentation is given, from Fig. 2 (c) it can be seen that passing through mixed Gaussian background modeling, area
Filtering and morphology post processing, sport foreground region is accurately estimated；In Fig. 2 (d), by by area mistake
Small connected region is filtered, and final result accurately returns the rectangle frame residing for all sport foreground objects in scene.
Fig. 3 is certain specified pedestrian's template image schematic diagram in this example, and Fig. 4 is that the present invention is returned for the template image
Retrieval result figure.As can be seen that 4 returning results are all the correct matchings to To Template image, even if personage's attitude
Change is there occurs with residing background, the present invention still can find correct result from complex scene.
All experiments are realized on common PC computers in this example, COMPUTER PARAMETER：Central processing unit Interl (R)
Core (TM) i3-2100 CPU@3.10GHz, internal memory 4GB.The processing speed of monitor video with scene moving object it is intensive
Degree is relevant, and the average processing speed per frame is 29ms in this example, adds frame-skipping operation and parallel processing, can be right in real time
Specified pedestrian or vehicle in monitoring network are used for quickly detecting identification.
Claims (5)
1. the quick detection recognition methods of the specified pedestrian or vehicle in a kind of video surveillance network, it is characterised in that including such as
Lower step：
Step 1：Taken from the picture comprising specified pedestrian or vehicle or video center by way of rectangle frame is demarcated first to be checked
Survey identification target；
Step 2：The To Template image taken to frame is normalized change of scale, and generates various mixing of To Template image
Feature；
Step 3：For the monitor video that each camera shoots opens up independent thread, each camera in monitoring network is located parallel
Reason, background modeling is carried out using gauss hybrid models respectively to the monitoring scene residing for each camera；
Step 4：The moving foreground object in each camera is extracted using area filtering and morphology post processing；Transported by setting
The moving-target area threshold moving target too small to filter area, to filter out other that pedestrian or vehicle are not belonging in scene
Moving object；
Step 5：Sport foreground is extracted to each monitor video frame-skipping, preliminary treatment requirement is met to all in video surveillance network
Moving object is normalized change of scale and calculates various composite characters, is weighted similitude with To Template image afterwards
Measurement；
Step 6：According to the similitude size with template image, the detected moving object in each monitor video in network will be monitored
Information be dynamically put into the result vector V of designated lengthresult, after all monitor videos are traveled through, return to the result vector conduct
Final detection recognition result.
2. the quick detection recognition methods of the specified pedestrian or vehicle in video surveillance network according to claim 1, its
It is characterised by：Various composite characters described in step 2 are main color characteristic, block margin direction histogram feature, piecemeal HSV
Histogram feature, HoG features and invariable rotary LBP features of equal value, its specific generating process is：
Step 2.1：The generating process of wherein domain color feature is：
1. the To Template image after Normalized Scale is converted is transformed into HSV space from rgb space, only extracts image in HSV
Chrominance component in space；
2. the span of chrominance component is divided into 8 intervals, by the chrominance component of To Template image according to this 8 intervals
Projection turns into the value chist in i-th interval of 8 dimension H histogram of component chist, wherein chistiObtained by following formula：
Wherein hx,yRefer to the chrominance component value of image (x, y) coordinate position pixel, Rect refers to To Template image or fortune
The foreground area that dynamic foreground detection goes out, δi(hx,y) be defined as follows：
3. the interval average tone value c of 8 tones is calculated respectivelyi, and 8 dimension histograms are normalized, obtainP is chosen from 8 dimension normalization histogramsiMaximum 3-dimensional, uses vector vdcPreserve the c corresponding to this 3-dimensionaliWith
piConstitute the domain color feature of the To Template image；Therefore domain color feature vdcIt is made up of 6 dimensions altogether, in having corresponded to target image
The tone value and percentage v of 3 maximum color components of accountingdc=[c1,p1,c2,p2,c3,p3]；
Step 2.2：The generating process of wherein block margin direction histogram is：
1. target image after Normalized Scale is converted first is converted into gray level image by RGB color image, and it is divided into 4 ×
4 totally 16 pieces；
2. gray level image is filtered using Sobel horizontal edges detective operators and Sobel vertical edge detective operators, is obtained
The horizontal edge intensity and vertical edge intensity of each pixel, and the edge direction e of each pixel is obtained using themx,y
And edge strength | ex,y|；
3. pixel is divided into five types according to edge direction and intensity：Horizontal edge, vertical edge, 45 ° of edges, 135 ° of sides
Edge and directionless edge；Edge orientation histogram is built using four kinds of edges above；
4. directionless edge threshold T is sete, in each piecemeal, T is more than to all edge strengthsePixel according to edge side
To statistics with histogram is carried out, edge orientation histogram is tieed up in generation 4；4 × 4 piecemeal symbiosis, will into 64 dimension edge orientation histograms
This 64 dimensional vector saves as local edge direction feature ehistl；4 dimension edge orientation histograms in 16 piecemeals are carried out again
It is cumulative, constitute the global edge direction characteristic ehist of 4 dimensions of view picture target imageg；They are merged into 68 dimensional vectors, are saved as
Block margin direction histogram feature ved；
Step 2.3：The wherein histogrammic generating process of piecemeal HSV is：
1. To Template image or sport foreground image after Normalized Scale is converted project to HSV colors by rgb color space
Color space, and level is divided into 4 pieces from top to bottom by image；
2. the HSV histograms in 4 piecemeals are calculated respectively, and wherein chrominance component is divided into 16 intervals, saturation degree component and brightness
Component is respectively divided into 4 intervals, and each piecemeal finally gives 24 dimension histogram features；4 24 dimension histograms are spliced to form view picture
96 dimension piecemeal HSV histograms of image；
Step 2.4：The generating process of wherein HoG features is：
1. the To Template image after Normalized Scale is converted is converted into gray level image, and carries out Gamma schools to gray level image
Just, the influence of local shades and illumination variation to characteristic extraction procedure is reduced；
2. convolution algorithm is done to image with [- 1,0,1] gradient operator, obtains the horizontal gradient G of each pixelx(x, y), then use
[1,0,-1]TGradient operator does convolution algorithm to image, obtains the vertical gradient G of each pixely(x,y)；And utilize horizontal ladder
Degree and vertical gradient calculate gradient magnitude and gradient direction：
3. the cell factory of multiple 8 × 8 pixels is divided the image into, gradient direction is averagely divided into 9 areas by span
Between, according to the gradient direction of each pixel in cell factory, histogram of gradients is tieed up in generation 9；
4. setting every 4 adjacent cell factory composition one piecemeal, i.e. each piecemeal has 16 × 16 pixels, and can phase between piecemeal
Mutually overlap；By 49 dimension histogram of gradients series connection of cell factory in each piecemeal, the piecemeal description vectors of 36 dimensions are obtained；
5. the window of setting 64 × 64, window is vertically slided from top to bottom, is slided at intervals of 8 pixels, by window
The description vectors of all piecemeals that mouth is included are connected, and constitute the description vectors of whole window, then retouching all windows
State vector to be connected, finally give HoG characteristic vectors vhog；
Step 2.5：The generating process of invariable rotary LBP features wherein of equal value is：
1. To Template image or sport foreground image after Normalized Scale is converted are converted into gray level image, every in figure
Individual pixel, compares itself and 8 gray value sizes of neighbor pixel around, and neighbours' point gray value is then set to more than the pixel
1, on the contrary then it is set to 0；8 numeric strings are unified into one 82 system numbers in the direction of the clock since 12 o'clock position again, are counted
Calculate formula as follows：
Wherein, P is neighbours' number, and R is radius size, gyAnd gcThe respectively gray value and center pixel of neighbor pixel point
Gray value；
2. each the 82 system numbers that will be obtained join end to end, formed an annular, by the circular clockwise rotate 7 times, altogether
8 groups 82 system numbers can be obtained, 8 group of 22 minimum system number of system number intermediate value is chosen, the number is the corresponding rotation of pixel
Turn constant LBP values；
3. the invariable rotary LBP values that will 2. be obtained by step are divided into two classes, and wherein 0-1 conversion times are not more than 2 times and are classified as one
Class, referred to as consistent LBP operators；Remaining 2 system number is all classified as another kind of；By above-mentioned classification, consistent invariable rotary LBP is calculated
Son has 9 kinds, and non-uniform LBP operators have a kind；
4. the fritter of multiple 16 × 16 pixels is divided the image into, pixel LBP values in each fritter are calculated, one 10 dimension is constituted
Invariable rotary LBP histograms of equal value；The LBP histograms of all fritters are connected, the rotation of equal value of 840 dimensions is finally given
Constant LBP features vLBP。
3. the quick detection recognition methods of the specified pedestrian or vehicle in video surveillance network according to claim 1, its
It is characterised by：The foreground target in each camera is extracted in the filtering of utilization area and morphology post processing described in step 4, its
Implement including following sub-step：
Step 4.1：Area filtering is carried out to the binary image after mixed Gaussian background modeling and thresholding treatment；
Step 4.2：3 expansive workings are carried out with 3 × 3 templates to binary image；
Step 4.3：1 etching operation is carried out with 3 × 3 templates again to binary image；
Step 4.4：Finally carry out an area filtering again to binary image；
Area filtering described in step 4 refers to the fritter that binary image is divided into multiple 4 × 4 pixels, if certain is small
When the number of foreground pixel point is less than or equal to 3 in block, then all pixels point in the fritter is defined as background pixel point, otherwise then
Retain the foreground pixel point in the fritter；
The moving target too small to filter area by setting moving target area threshold described in step 4, it is implemented
Process is：All of connected region in binary image is found out after treatment, the number of each connected region foreground pixel point is calculated,
Setting area threshold ThcIt is the 1/400 of video frame images area, if the area of connected region is more than ThcThen retain connected region,
And the minimum rectangle frame comprising the connected region is returned as sport foreground region；If connected region area is less than or equal to ThcThen
The connected region is set to background area.
4. the quick detection recognition methods of the specified pedestrian or vehicle in video surveillance network according to claim 1, its
It is characterised by：Sport foreground is extracted to each monitor video frame-skipping described in step 5, it implements process and is：According to user
, one by one be weighted the sport foreground object extracted in step 4 with To Template image again every F frames by the frame-skipping number F of setting
Similarity measurement；
Weighting similarity measurement described in step 5, it implements process and is：Sport foreground object and target are first calculated respectively
The domain color characteristic distance D of template imagedc, block margin direction histogram characteristic distance Ded, piecemeal HSV histogram features away from
From Dhsv, HoG characteristic distances DhogWith invariable rotary LBP characteristic distances D of equal valueLBP, then the distance of five kinds of features is normalized
Weighting similarity measurement, obtains weighting similarity distance Dall：
Wherein α, β, γ, λ,It is normalized weighing factors, for balancing influence of each feature to overall similarity measurement, according to
Specific environment is tested to adjust in advance.
5. the quick detection recognition methods of the specified pedestrian or vehicle in video surveillance network according to claim 1, its
It is characterised by：The information that will monitor the detected moving object in network in each monitor video described in step 6 is dynamically put into finger
The result vector V of measured lengthresult, it is implemented including following sub-step：
Step 6.1：Create a vectorial V with designated length Nresult, wherein N be need return monitor video network in
With specified pedestrian or the immediate object number of car modal image；
Step 6.2：When there is new sport foreground to be detected in monitor video network, V is first determined whetherresultThe object of middle preservation
Number；
Step 6.3：If VresultThe object number of middle preservation is less than N, when object and the weighting similarity distance of To Template image
DallMeet similarity threshold Th set in advanceallWhen, then the residing camera of the object numbered, frame number occur, square occur
The information such as the subject image after shape frame coordinate position in the frame and Normalized Scale conversion are retained in vectorial VresultIn；
Step 6.4：If VresultThe object number of middle preservation has reached N, and when object and the weighting similitude of To Template image
Apart from DallMeet similarity threshold Th set in advanceallWhen, then the weighting similarity distance D for being obtained with the objectallWith
VresultIn the weighting similarity distance that obtains of the object maximum with To Template distanceCompare, ifThen
V is substituted with the objectresultIn the object maximum with To Template distance, and by VresultN number of object of middle preservation according to mesh
Mark template similarity distance is resequenced from small to large；
Step 6.5：After the monitor video that all cameras shoot in traversal monitoring network, the V of returnresultIt is in store in vector
With specified pedestrian or the immediate N number of object of car modal image in whole monitoring network, and comprising taking the photograph residing for each of which
As head numbering, the frame number for occurring, there is rectangle frame coordinate position in the frame and Normalized Scale conversion after object figure
As information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410356465.0A CN104166841B (en) | 2014-07-24 | 2014-07-24 | The quick detection recognition methods of pedestrian or vehicle is specified in a kind of video surveillance network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410356465.0A CN104166841B (en) | 2014-07-24 | 2014-07-24 | The quick detection recognition methods of pedestrian or vehicle is specified in a kind of video surveillance network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104166841A CN104166841A (en) | 2014-11-26 |
| CN104166841B true CN104166841B (en) | 2017-06-23 |
Family
ID=51910645
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410356465.0A Active CN104166841B (en) | 2014-07-24 | 2014-07-24 | The quick detection recognition methods of pedestrian or vehicle is specified in a kind of video surveillance network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104166841B (en) |
Families Citing this family (43)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104463151A (en) * | 2015-01-05 | 2015-03-25 | 北京邮电大学 | Rapid target matching method based on fusion of main color histograms and spatial position information |
| ES2744924T3 (en) * | 2015-01-14 | 2020-02-26 | Saronikos Trading & Services Unipessoal Lda | Apparatus and procedure for detecting entities in an image |
| CN104796664A (en) * | 2015-03-26 | 2015-07-22 | 成都市斯达鑫辉视讯科技有限公司 | Video monitoring device |
| CN106295466A (en) * | 2015-05-18 | 2017-01-04 | 佳能株式会社 | Image processing method and device |
| CN104866616B (en) * | 2015-06-07 | 2019-01-22 | 中科院成都信息技术股份有限公司 | Monitor video Target Searching Method |
| CN105005773A (en) * | 2015-07-24 | 2015-10-28 | 成都市高博汇科信息科技有限公司 | Pedestrian detection method with integration of time domain information and spatial domain information |
| CN106559645B (en) * | 2015-09-25 | 2020-01-17 | 杭州海康威视数字技术股份有限公司 | Monitoring method, system and device based on camera |
| CN105447448B (en) * | 2015-11-12 | 2019-01-11 | 江南大学 | A kind of bayonet vehicle localization method based on Gauss color distance |
| CN105447457A (en) * | 2015-11-15 | 2016-03-30 | 安徽清新互联信息科技有限公司 | License plate character identification method based on adaptive characteristic |
| CN105869148A (en) * | 2016-03-24 | 2016-08-17 | 北京小米移动软件有限公司 | Target detection method and device |
| CN105957093B (en) * | 2016-06-07 | 2018-12-25 | 浙江树人大学 | A kind of texture distinguishes the ATM delay object detecting method of optimization HOG operator |
| JP6598741B2 (en) * | 2016-07-21 | 2019-10-30 | 三菱電機ビルテクノサービス株式会社 | Image processing apparatus and program |
| CN106504264B (en) * | 2016-10-27 | 2019-09-20 | 锐捷网络股份有限公司 | Video foreground image extraction method and device |
| CN106845359A (en) * | 2016-12-26 | 2017-06-13 | 四川农业大学 | Tunnel portal driving prompt apparatus and method based on infrared emission |
| CN107239780A (en) * | 2017-04-29 | 2017-10-10 | 安徽慧视金瞳科技有限公司 | A kind of image matching method of multiple features fusion |
| CN107292892B (en) * | 2017-06-23 | 2020-04-21 | 北京奇艺世纪科技有限公司 | Video frame image segmentation method and device |
| CN107704797B (en) * | 2017-08-08 | 2020-06-23 | 深圳市安软慧视科技有限公司 | Real-time detection method, system and equipment based on pedestrians and vehicles in security video |
| CN107578045A (en) * | 2017-09-19 | 2018-01-12 | 北京工业大学 | A kind of Underwater targets recognition based on machine vision |
| CN107918767B (en) * | 2017-11-27 | 2019-11-08 | 北京旷视科技有限公司 | Object detection method, device, electronic equipment and computer-readable medium |
| CN107895376A (en) * | 2017-12-11 | 2018-04-10 | 福州大学 | Based on the solar panel recognition methods for improving Canny operators and contour area threshold value |
| CN110110111B (en) * | 2018-02-02 | 2021-12-31 | 兴业数字金融服务（上海）股份有限公司 | Method and device for monitoring screen |
| CN108921083B (en) * | 2018-06-28 | 2021-07-27 | 浙江工业大学 | Illegal mobile vendor identification method based on deep learning target detection |
| CN109389031B (en) * | 2018-08-27 | 2021-12-03 | 浙江大丰实业股份有限公司 | Automatic positioning mechanism for performance personnel |
| CN109214996B (en) * | 2018-08-29 | 2021-11-12 | 深圳市元征科技股份有限公司 | Image processing method and device |
| CN110529788A (en) * | 2018-10-30 | 2019-12-03 | 永康市道可道科技有限公司 | The explosion-proof pendent lamp of automatic projection formula |
| BR112021004557B1 (en) | 2018-11-09 | 2023-09-26 | Beijing DIDI Infinity Technology and Development Co., Ltd | SYSTEM AND METHOD FOR DETECTING CONFLICTS IN VEHICLES |
| CN109635720A (en) * | 2018-12-10 | 2019-04-16 | 江南大学 | The illegal road occupying real-time detection method actively monitored based on video |
| CN110232314A (en) * | 2019-04-28 | 2019-09-13 | 广东工业大学 | A kind of image pedestrian's detection method based on improved Hog feature combination neural network |
| CN110263788B (en) * | 2019-07-17 | 2021-07-30 | 广西盖德科技有限公司 | Method and system for quickly identifying vehicle passing |
| CN110378341A (en) * | 2019-07-24 | 2019-10-25 | 西南交通大学 | A kind of binocular vision pedestrian distance detection method |
| CN110517183B (en) * | 2019-07-26 | 2020-09-25 | 电子科技大学 | High-speed low-power-consumption image processor based on retina mechanism |
| CN110490253A (en) * | 2019-08-19 | 2019-11-22 | 浙江工业大学 | A kind of image similarity calculation method based on spatial color EMD |
| CN110516649B (en) * | 2019-09-02 | 2023-08-22 | 南京微小宝信息技术有限公司 | Face recognition-based alumni authentication method and system |
| CN110765926B (en) * | 2019-10-21 | 2023-05-09 | 杭州易现先进科技有限公司 | Picture book identification method, device, electronic equipment and storage medium |
| CN110956200A (en) * | 2019-11-05 | 2020-04-03 | 哈尔滨工程大学 | Tire pattern similarity detection method |
| CN111062971B (en) * | 2019-12-13 | 2023-09-19 | 深圳龙岗智能视听研究院 | Deep learning multi-mode-based mud head vehicle tracking method crossing cameras |
| CN111179302B (en) * | 2019-12-27 | 2023-03-24 | 普联国际有限公司 | Moving target detection method and device, storage medium and terminal equipment |
| CN111461242A (en) * | 2020-04-08 | 2020-07-28 | 北京航天新风机械设备有限责任公司 | Multi-material rapid comparison and matching method for production line |
| CN112149581A (en) * | 2020-09-25 | 2020-12-29 | 北京嘀嘀无限科技发展有限公司 | Working condition monitoring method, system, device, electronic equipment and readable storage medium |
| CN112232153A (en) * | 2020-09-30 | 2021-01-15 | 广东职业技术学院 | Method and system for acquiring track of target person |
| CN113255705B (en) * | 2021-07-15 | 2021-11-05 | 广东省大湾区集成电路与系统应用研究院 | Intelligent identification device for protective color animal |
| CN113807328B (en) * | 2021-11-18 | 2022-03-18 | 济南和普威视光电技术有限公司 | Target detection method, device and medium based on algorithm fusion |
| CN117079219B (en) * | 2023-10-08 | 2024-01-09 | 山东车拖车网络科技有限公司 | Vehicle running condition monitoring method and device applied to trailer service |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102012213A (en) * | 2010-08-31 | 2011-04-13 | 吉林大学 | Method for measuring foreground height through single image |
| CN102592454A (en) * | 2012-02-29 | 2012-07-18 | 北京航空航天大学 | Intersection vehicle movement parameter measuring method based on detection of vehicle side face and road intersection line |
| CN102708367A (en) * | 2012-03-11 | 2012-10-03 | 无锡派图半导体设备有限公司 | Image identification method based on target contour features |
| CN103049751A (en) * | 2013-01-24 | 2013-04-17 | 苏州大学 | Improved weighting region matching high-altitude video pedestrian recognizing method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20100065677A (en) * | 2008-12-08 | 2010-06-17 | 한국전자통신연구원 | Method for detection of multi moving objects in the high resolution image sequences and system thereof |
-
2014
- 2014-07-24 CN CN201410356465.0A patent/CN104166841B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102012213A (en) * | 2010-08-31 | 2011-04-13 | 吉林大学 | Method for measuring foreground height through single image |
| CN102592454A (en) * | 2012-02-29 | 2012-07-18 | 北京航空航天大学 | Intersection vehicle movement parameter measuring method based on detection of vehicle side face and road intersection line |
| CN102708367A (en) * | 2012-03-11 | 2012-10-03 | 无锡派图半导体设备有限公司 | Image identification method based on target contour features |
| CN103049751A (en) * | 2013-01-24 | 2013-04-17 | 苏州大学 | Improved weighting region matching high-altitude video pedestrian recognizing method |
Non-Patent Citations (1)
| Title |
|---|
| 基于高斯混合建模的多尺度HOG行人头肩特征检测;芮挺 等;《山东科技大学学报(自然科学版)》;20130430;第32卷(第2期);第90-93页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104166841A (en) | 2014-11-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104166841B (en) | The quick detection recognition methods of pedestrian or vehicle is specified in a kind of video surveillance network | |
| CN105518709B (en) | The method, system and computer program product of face for identification | |
| Luo et al. | Traffic sign recognition using a multi-task convolutional neural network | |
| CN103942577B (en) | Based on the personal identification method for establishing sample database and composite character certainly in video monitoring | |
| CN108875600A (en) | A kind of information of vehicles detection and tracking method, apparatus and computer storage medium based on YOLO | |
| CN103824070B (en) | A kind of rapid pedestrian detection method based on computer vision | |
| Rachmadi et al. | Vehicle color recognition using convolutional neural network | |
| Wang et al. | Detection and localization of image forgeries using improved mask regional convolutional neural network | |
| CN109819208A (en) | A kind of dense population security monitoring management method based on artificial intelligence dynamic monitoring | |
| CN105404886B (en) | Characteristic model generation method and characteristic model generating means | |
| CN109447169A (en) | The training method of image processing method and its model, device and electronic system | |
| CN105574550A (en) | Vehicle identification method and device | |
| CN106372666B (en) | A kind of target identification method and device | |
| CN106096602A (en) | A kind of Chinese licence plate recognition method based on convolutional neural networks | |
| CN107273832B (en) | License plate recognition method and system based on integral channel characteristics and convolutional neural network | |
| CN109815874A (en) | A kind of personnel identity recognition methods, device, equipment and readable storage medium storing program for executing | |
| CN109190475B (en) | Face recognition network and pedestrian re-recognition network collaborative training method | |
| CN105893946A (en) | Front face image detection method | |
| CN106295600A (en) | Driver status real-time detection method and device | |
| CN107085696A (en) | A kind of vehicle location and type identifier method based on bayonet socket image | |
| CN106610969A (en) | Multimodal information-based video content auditing system and method | |
| CN104504395A (en) | Method and system for achieving classification of pedestrians and vehicles based on neural network | |
| CN107730515A (en) | Panoramic picture conspicuousness detection method with eye movement model is increased based on region | |
| CN109711322A (en) | A kind of people's vehicle separation method based on RFCN | |
| CN112560675B (en) | Bird visual target detection method combining YOLO and rotation-fusion strategy |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |