CN103034852B - The detection method of particular color pedestrian under Still Camera scene - Google Patents
The detection method of particular color pedestrian under Still Camera scene Download PDFInfo
- Publication number
- CN103034852B CN103034852B CN201210586125.8A CN201210586125A CN103034852B CN 103034852 B CN103034852 B CN 103034852B CN 201210586125 A CN201210586125 A CN 201210586125A CN 103034852 B CN103034852 B CN 103034852B
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- foreground
- color
- histogram
- foreground blocks
- 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.)
- Expired - Fee Related
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides the detection method of particular color pedestrian under a kind of Still Camera scene, the method adopts gauss hybrid models to carry out modeling to the background of scene, by obtaining rough sport foreground with the comparison when pre-treatment image after obtaining background model, by sport foreground being proposed with the form of rectangular block after filtering and Morphological scale-space, be referred to as foreground blocks; Foreground point proportion in each foreground blocks calculated and distinguishes complicated foreground blocks and simple prospect block in conjunction with foreground point statistical property of projection number in horizontal coordinate, and estimating the horizontal coordinate that may there is pedestrian; For complicated foreground blocks, utilize online pedestrian level estimation model and based on the Adaboost sorter of histograms of oriented gradients feature, pedestrian target accurately detected; After obtaining pedestrian target position, its clothing color is identified.The present invention improves detection speed and the accuracy of pedestrian target, can be applicable in real-time video monitoring and video frequency searching.
Description
Technical field
What the present invention relates to is a kind of method of Computer Vision and mode identification technology, specifically a kind of particular color pedestrian detection method being applied to video monitoring system.
Background technology
The main task that pedestrian target detects is the object marking to realize retrieval to the pedestrian in video sequence.The detection of particular color pedestrian is then on the basis of pedestrian's target detection, identify the pedestrian target with particular color further, such as, in given video, detect the pedestrian wearing blue coat.The detection technique of particular color pedestrian can be applicable to the field such as video monitoring, video frequency searching, thus the burden that the monitor video data reducing magnanimity are brought to operating personnel.
Typical pedestrian detection method is based on background modeling, compare according to current video pictorial information and background model, extract foreground blocks and each foreground blocks is all used as independently object, then directly the feature of Utilization prospects block carries out object classification identification, the people such as such as J.Renno exist " Object Classification in VisualSurveillance Using Adaboost " in the method mentioned, this papers included was at " ComputerVision and Pattern Recognition " meeting collection the 1 to 8 page in 2007.The advantage of this method is to find foreground blocks fast according to background model, and identifying is simple, therefore can reach good real-time.But many times may occur multiple object in a foreground blocks, because mutually blocking may appear in moving target, in the extraction process of prospect, they can be treated as a foreground blocks, and whole foreground blocks just may be judged to be an object by the method.
For the pedestrian detection problem in static images, the people such as Navneet Dalal propose the shape of utilization orientation histogram of gradients feature HOG to pedestrian and learn and train SVM classifier in " ComputerVision and Pattern Recognition " meeting in 2005, go by sorter the method differentiating and detect pedestrian target.This papers included is at the 886 to 893 page of " Computer Vision and Pattern Recognition " meeting collection in 2005.Because the shape of different pedestrian is all more similar, and this method allows the limb action that pedestrian has some trickle, these trickle limb actions can be left in the basket and not affect Detection results, and this method is widely used and develops in the pedestrian detection of static images.But this method will consume for a long time in feature extraction and assorting process, it will be made to be difficult to be applied to field of video monitoring separately.
In Chinese patent " pedestrian detection method based on Analysis on Prospect and pattern-recognition " (publication No. is CN102147869A), the method that this patent combines background modeling and detection of classifier detects pedestrian, the method and pedestrian detection method framework of the present invention there is similarity, foreground information and sorter are all used, but in the method, the Height Estimation of pedestrian is relied on to the pedestrian level model of priori, the manual height carrying out people is needed to demarcate before detection, and the position of the method for profile peak point estimation pedestrian of looking for described in this patent may consume the more time.The present invention is directed to above problem and all done solution, adopt adaptive pedestrian level model to eliminate manual process of demarcating, the statistical nature of Utilization prospects point instead of contour feature make the estimating speed of pedestrian position be improved.
The main task of color detection identifies the body color of interesting target, concrete application is had in fields such as vehicle detection, such as Mengjie Yang refer to a kind of method detecting vehicle color in " Vehicle Color Recognition Using Monocularcamera ", and this papers included is at the 1 to the 5 page of " WirelessCommunication andSignal Processing " meeting collection in 2011.The method realizing color detection has a lot, such as Brown, L.M. a kind of method for detecting color based on HSV space segmentation mentioned in " Color Retrieval for Video Surveillance ", this papers included is at the 283 to 290 page of " Advanced Video andSignal Based Surveillance " meeting collection in 2008.The method advantage is to be the detection to the main color of a few class by color detection problem reduction, and image processes in hsv color space, reduces the interference of illumination to color detection to a certain extent.
Summary of the invention
For defect of the prior art, the object of this invention is to provide a kind of particular color pedestrian detection method based on sport foreground analysis and static images detection technique of improvement, both speed and the accuracy of pedestrian detection had been improved, the clothing color of pedestrian can be identified again on the basis of pedestrian detection, improve practicality.
According to an aspect of the present invention, provide a kind of method detecting pedestrian target in video, the method comprises the steps:
The first step, gauss hybrid models is adopted to carry out modeling to the background of scene, by obtaining rough sport foreground with the comparison when pre-treatment image after obtaining background model, by sport foreground being proposed with the form of rectangular block after filtering and Morphological scale-space, be referred to as foreground blocks;
Second step, calculates foreground point proportion in each foreground blocks and distinguishes complicated foreground blocks and simple prospect block in conjunction with foreground point statistical property of projection number in horizontal coordinate, and estimating the horizontal coordinate that may there is pedestrian;
3rd step, determines whether pedestrian for simple prospect block by its size; For complicated foreground blocks, utilize online pedestrian level estimation model and based on the Adaboost sorter of histograms of oriented gradients feature, pedestrian target accurately detected, obtaining the pedestrian target in video.
Preferably, the horizontal coordinate method of estimation of described pedestrian is as follows: the foreground point utilizing gauss hybrid models to obtain projects in horizontal coordinate, the foreground point number that statistics drops in each horizontal coordinate generates histogram, to the smoothing filtering of the histogram obtained, the position that there is convex closure after filtering in histogram is exactly the position that may have pedestrian, just can estimate the horizontal level that may there is pedestrian by carrying out convex closure detection to histogram.
The projection histogram computing formula of described foreground point in horizontal coordinate is as follows
Wherein Histogram (x) represents the value of a histogrammic xth passage, x corresponds to the horizontal ordinate of image, histogrammic port number equals the width of image, height corresponds to the height of image, I (x, y) corresponding to the value of two-value foreground image at coordinate (x, y);
Preferably, described simple prospect block refers to the foreground blocks comprising single body, complicated foreground blocks refers to the foreground blocks comprising multiple object, the mode distinguishing simple prospect block and complicated foreground blocks is as follows: for simple prospect block, the ratio of foreground blocks shared by its foreground point pixel usually higher (being such as usually more than or equal to 0.45), whether may be simple prospect block by the threshold value preliminary judgement handling object of foregrounding point ratio, if it is determined that be simple prospect block, the method that recycling pedestrian horizontal coordinate is estimated checks wherein whether may have more than one pedestrian target, wherein may comprise multiple pedestrian target if estimated, foreground blocks is grouped in complicated foreground blocks, otherwise be simple prospect block.Distinguishing a principle that is simple and complicated foreground blocks is: simple prospect block can be divided in complicated foreground blocks, but not allow complicated foreground blocks to incorporate in simple prospect block.
Preferably, described online pedestrian level estimation model obtains in the following manner: assuming that the height Gaussian distributed of different pedestrians in scene, for its height in scene of same pedestrian with the change of vertical coordinate linear change, thus can with a conditional Gaussian function p (h|y about vertical coordinate y, pedestrian level in scene β) is described, this model is not used to remove to estimate the height of pedestrian in the early stage, training Height Estimation model is removed as training sample with the height of the pedestrian target being judged to obtain by simple prospect block and coordinate information, p (h|y is recycled after being no less than 30 training, β) pedestrian level of diverse location in scene is estimated.
Pedestrian level estimation model is
Wherein h (y) is the height of pedestrian target when ordinate is y, obeys the Gaussian distribution about condition y, and β is the parameter to be asked of model, σ
2for the distribution variance of specifying.
In the method for line computation β based on maximum the feel relieved estimation technique and Robbins-Monro algorithm, formula is as follows:
Wherein β
(N)be the model parameter obtained after the N time training, H
nbe the height of N number of pedestrian's sample, Y
nbe the y coordinate of N number of pedestrian's sample, N ' is stable factor, δ
2for the distribution variance of specifying.
According to a further aspect of the invention, provide the detection method of particular color pedestrian under a kind of Still Camera scene, described method in two steps, first adopts the above-mentioned first step to detect to the 3rd step and pedestrian target in video then identifies to its clothing color.
Preferably, described to its clothing color identify, specifically: after obtaining pedestrian target, utilize half body parted pattern by the upper body of pedestrian and the lower part of the body separated; Utilize method identification pedestrian's jacket of HSV space statistical color histogram and the color of trousers, finally mark the pedestrian target meeting setpoint color.
Preferably, in described half body parted pattern, calculate gradient map, specifically: for the sport foreground point of pedestrian, in RGB tri-passages, calculate the gradient of vertical direction respectively according to Soble operator, after summation, obtain total gradient map; Obtain the Grad cumulative sum of often row pixel accordingly further, obtain the transverse projection of gradient.When simple clothing, clothing has obvious border up and down, and this model obtains the gradient map of reflection color change according to the RGB component of the corresponding boxed area of pedestrian target, then does transverse projection analysis to gradient map, just can obtain lower part of the body cut-off rule.In order to guarantee the accuracy split further, limiting cut-off rule and can only drop in centre 1/3 scope of foreground blocks vertical direction.
Preferably, described HSV space statistical color histogram method comprises HSV space segmentation and statistical color histogram two parts, the former refers to and HSV space is divided into Red, yellow, green, blue, purple, black, white, ash 8 color regions, the latter refers to the color conversion of carrying out from rgb space to HSV space to image, and statistics drops on the pixel number of above-mentioned 8 color regions, and then obtain the histogram reflecting color distribution, analyze histogram and can draw clothing color.
Described HSV space segmentation, refers to, the piecewise linear function according to V component and S component distinguishes colored and achromatic area; For achromatic area, determined the cut-off rule of black-white-gray 3 color regions by V component; For colored region, carried out the segmentation of Red, yellow, green, blue, purple 5 color regions by H component.
Said method principle of the present invention is as follows: the region of Utilization prospects detection method determination moving target, the complicacy of each foreground area is judged, the foreground area that may comprise single body carries out Direct Recognition according to its yardstick, to the complicated foreground area comprising multiple object, utilize the horizontal level of statistical property quick position each pedestrian of pedestrian foreground point in horizontal coordinate, near the pedestrian position of prediction, utilize the Adaboost sorter based on histograms of oriented gradients feature accurately to detect pedestrian target, both foreground information is taken full advantage of like this, eliminate the region of unnecessary detection, improve pedestrian detection speed, the accuracy of pedestrian detection is improve again by the method for pattern-recognition, carry out to pedestrian human body segmentation's knowledge that half body segmentation is then color distortion and the priori relying on the upper lower part of the body of pedestrian to exist, the identification of color then make use of the cutting techniques of HSV space.
Compared with prior art, the present invention has following beneficial effect:
The present invention is on the basis that improve pedestrian target detection speed and accuracy, also achieve the identification of the clothing color to pedestrian, the present invention not only can be applied in traditional video monitoring system, realize the detection of pedestrian target, and the pedestrian target that particular color is worn clothes can also be detected in video frequency searching.
Accompanying drawing explanation
By reading the detailed description done non-limiting example with reference to the following drawings, other features, objects and advantages of the present invention will become more obvious:
Fig. 1 is general frame schematic diagram of the present invention.
Fig. 2 is that foreground extraction and pedestrian's horizontal coordinate estimate schematic diagram.
Fig. 3 is the key diagram that pedestrian and vehicle foreground overlap.
Fig. 4 is HSV space segmentation schematic diagram.
Fig. 5-Fig. 7 is the schematic diagram realizing particular color pedestrian detection.
Embodiment
Below in conjunction with specific embodiment, the present invention is described in detail.Following examples will contribute to those skilled in the art and understand the present invention further, but not limit the present invention in any form.It should be pointed out that to those skilled in the art, without departing from the inventive concept of the premise, some distortion and improvement can also be made.These all belong to protection scope of the present invention.
Embodiment
The present embodiment has mainly been developed under VS2008 platform, kernel program C++ writes, image procossing has used OpenCV1.0, and human-computer interaction interface VC++ develops, and the training program Matlab based on the Adaboost sorter of histograms of oriented gradients feature writes.Used by test is a video (768 × 576 pixels, 25fps) about multirow people walking, has a large amount of pedestrians in walking in scene, has the phenomenons such as parallel and overlap to occur.
The specific implementation step of the present embodiment is as follows:
The first step, extracts the histograms of oriented gradients feature HOG of pedestrian's image and non-pedestrian image in pedestrian's picture library, and the HOG characteristic procedure extracting single image is as follows:
(1) because HOG feature does not consider the color characteristic of picture, first RGB image is converted into gray level image.
(2) Grad at each some place in entire image is calculated: set up a gradient map in the X direction and a gradient map in the Y direction respectively, adopt discrete differential template during compute gradient figure, template is [-1 in the X direction, 0,1], template is [-1 in the Y direction, 0,1]
t.To obtain in image X-direction, after gradient map and Y-direction gradient map, the actual gradient direction of this pixel and the amplitude of its corresponding gradient can being calculated.
(3) ask the histogram of gradients of cell factory in image: the rectangular pixel area with 8 × 8 represents a cell factory, every 4 cell factory form a basic block, first obtain the histogram of gradients of each cell factory.The histogram that undirected gradient 0-180 and 9 histogram passage describes cell factory is have employed in embodiment, histogrammic process of establishing can be regarded as each pixel in cell factory and according to its gradient, each passage is weighted to the process of ballot, wherein weights gradient amplitude represents, such as the gradient direction of certain point may be 18, its gradient amplitude is 2, then should be passage 1 and add numerical value 2.
(4) 4 cell factory are spliced into a rectangle block, because each cell factory has 9 passages, then the histogram feature of a rectangle block has 36 passages.Preliminary splicing utilizes following formula feature to be processed after obtaining the histogram feature of rectangle block:
Wherein v is untreated histogram feature vector, and v ' is the proper vector after process,
for the second order norm of v, ∈ be one prevent denominator be zero minimum constant.
Again each component is being normalized after above formula process, thus is obtaining more stable block histogram feature.
(5) choose with the interval of 8 pixels the histograms of oriented gradients that overlapping block carries out being spliced to form a Description Image in the picture, because the training sample picture yardstick used in embodiment is 128 × 64,105 rectangle blocks are obtained so can choose altogether, each block has 36 dimensions, the intrinsic dimensionality then describing pedestrian's sample is 3780, i.e. the dimension of a HOG feature.
Second step, by the HOG feature obtained from pedestrian's image and non-pedestrian image zooming-out as positive and negative training sample set χ, obtains sorter with the training of Adaboost method.
χ＝{<x
1，y
1>，<x
2，y
2>…<x
L，y
L>|x
i∈R
3780，y
i∈{1，-1}}，
Wherein x
ibe the feature of i-th sample, y
ibe the mark of i-th sample, y
i=1 represents that i-th sample is pedestrian's sample, y
i=-1 represents that i-th sample is non-pedestrian sample.
3rd step, utilizes gauss hybrid models to carry out modeling to the background of scene.
4th step, by obtaining preliminary foreground image when pre-treatment video image and background model comparison, carries out filtering and Morphological scale-space to remove noise spot and to fill up little hole to foreground image.Profile block in the prospect OpenCV obtained after process is extracted function moving target is extracted with the form of rectangular block.
5th step, calculate the ratio of foreground blocks shared by foreground point in each foreground blocks, threshold value T=0.45 is set, when the foreground point proportion of foreground blocks is less than T then by this foreground blocks Direct Classification in complicated foreground blocks, otherwise this foreground blocks preliminary judgement is simple prospect block and accepts to judge further.
6th step, carries out the estimation of pedestrian's horizontal level to each foreground blocks.Method of estimation (be described with the method for estimation of entire image, carry out such pedestrian's horizontal coordinate estimate during application to each foreground blocks) is as follows:
First the projection histogram of foreground point in X-coordinate is calculated,
Wherein Histogram (x) represents the value of a histogrammic xth passage, and x corresponds to the horizontal ordinate of image, and height corresponds to the height of image, and I (x, y) is corresponding to the value of two-value foreground image at coordinate (x, y).
After obtaining the projection histogram of foreground point in X-coordinate, by the smooth template of [0.12,0.18,0.4,0.18,0.12], filtering is carried out to histogram.
Look for the peak point of convex closure in histogram after the filtering, its horizontal coordinate is as the pedestrian's horizontal coordinate estimated.
7th step, the estimated result of the simple prospect block tentatively obtained in 5th step in the 6th step is verified further, if estimate in preliminary foreground blocks that the pedestrian target obtained is more than 1, again incorporates this foreground blocks into complicated foreground blocks, otherwise confirms as simple prospect block.
8th step, to the simple prospect block obtained, directly carries out target discrimination according to its length breadth ratio and yardstick, and when the length breadth ratio of foreground blocks is in scope [1.6,3], and its area is less than video image area
time, then assert that this foreground blocks is pedestrian target.
9th step, is judged that by simple prospect block the height of the pedestrian target obtained and Y-coordinate value (H, Y) train Height Estimation model as training sample
wherein h (y) is height when pedestrian target coordinate is y, obeys the Gaussian distribution about condition y, and β is the parameter to be asked of model, σ
2for the distribution variance of specifying.Train and be iteratively:
Wherein β
(N)be the model parameter obtained after the N time training, H
nbe the height of N number of pedestrian's sample, Y
nbe the y coordinate of N number of pedestrian's sample, the factor of N '=100 for making model fast and stable add.
Tenth step, carries out regional area window Scanning Detction pedestrian to the Adaboost sorter that each complicated foreground blocks second step obtains.
The method calculating local detection area is as follows:
First estimate the maximum pedestrian level that may occur in foreground blocks, the bottom Y-coordinate y ' by foreground blocks substitutes into pedestrian level estimation model and asks its conditional mean, obtains h
max=E (p (h) | y ').
Then maximum pedestrian's width w that may occur is calculated
max=0.5h
max+ 5.
In the 6th step, obtain the X-coordinate position that may occur pedestrian of estimation, centered by these coordinates, left and right respectively extends 0.5w
maxit is exactly each surveyed area scope in the horizontal direction.
Specify that the scope of surveyed area is identical with each foreground blocks scope in the vertical direction in vertical direction.
The detection yardstick used during detection is determined according to the pedestrian level estimated, the Gaussian distribution that h (y) obeys is carried out stochastic sampling 2 pedestrian levels, add E (p (h) | y) totally 3 height, change the yardstick of surveyed area according to these 3 height values, then carry out the pedestrian in window Scanning Detction region.
11 step, carry out half body segmentation to the pedestrian detected, dividing method is as follows:
Calculate the gradient map of the corresponding foreground point of pedestrian target,
gradient (i, j) represents the pixel gradient value of the i-th row jth row, C
k(i, j) represents a kth Color Channel of the i-th row jth row pixel, and Soble represents the operator asking gradient.
Carry out transverse projection according to gradient map, obtain
Sum (i)=∑
jgradient (i, j), Sum (i) represent the i-th row pixel gradient and, traversal pedestrian target in the middle of 1/3 row pixel, find the i making Sum (i) maximum, in this, as the Y-coordinate that cut-off rule is corresponding.
12 step, carry out colour recognition to the upper lower part of the body of pedestrian, method is as follows:
Color space conversion is carried out to upper lower part of the body image, obtains the data of HSV space.Piecewise linear function according to V component and S component distinguishes colored and achromatic area, for achromatic area, determined the cut-off rule of black-white-gray 3 color regions by V component, for colored region, carried out the segmentation of Red, yellow, green, blue, purple 5 color regions by H component.Count the number percent that the foreground point number dropping on 8 color regions accounts for foreground point sum respectively, and then filter out the heavy maximum color of accounting in foreground point, in this, as the color of clothing.
Implementation result and explanation
According to method above, test is done to the video sequence of one section of pedestrian's walking.Fig. 2 gives the extraction of foreground blocks and the process of the counting estimation pedestrian horizontal level that projected in X-coordinate by foreground point.The figure on the first row left side is the video original image when pre-treatment, the figure on the second row left side is the sport foreground image obtained by gauss hybrid models, white box on the right of the first row is the foreground blocks obtained by the contours extract function in OpenCV, the image on the third line left side is that foreground point projects the histogram that counting generates in X-coordinate, image on the right of second row is the projection histogram of foreground point in X-coordinate after smoothing processing, and the image on the right of the third line is pedestrian position estimation effect figure.Can find out that the convex closure of foreground point on X-coordinate in projection histogram can correspond to a pedestrian substantially from the effect provided, but the convex closure number that may be formed between the pedestrian of parallel walking sometimes more than effective strength, the pedestrian's horizontal level in this way obtained is only the position estimated, reduces the scope of detection.
Fig. 3 is the key diagram carrying out the estimation of pedestrian's horizontal coordinate when generation prospect is overlapping, because pedestrian's horizontal coordinate of the present invention is estimated to detect based on foreground pixel point fast, when occurring the situation having vehicle after pedestrian, if camera horizontal positioned, the prospect that possible vehicle is formed can cover pedestrian target and cause pedestrian detection failure.The present invention is applied to monitoring scene, general camera is all downward-sloping placement, in this case, the knee of pedestrian usually with lower part can be reflected in the foreground instead of be covered by object after one's death, as in Fig. 3, left side gray area represents vehicle, left side darker regions represents that pedestrian can not by the part of vehicle covering below, this part still can form little convex closure after X-coordinate carries out projection statistics, so the pedestrian's horizontal level method of estimation used in invention substantially can estimate the pedestrian of all existence in monitoring scene.
Fig. 4 is the segmentation schematic diagram to hsv color space, rhombus represents V, S plane, middle disk represents H, S plane, and two lines of approximate Double curve represent the cut-off rule of colored region and the achromatic area determined in H, S plane (being similar to piecewise linear function in reality).
Fig. 5,6,7 is the design sketchs detecting white jacket pedestrian in test video, although can find out that pedestrian in scene is a lot of and pedestrian has the situation of mutually blocking to occur each other, but the program of embodiment can arrive all pedestrian detection substantially, and accurately identify the pedestrian wearing white jacket.
All experiments all realize on PC computing machine, and the parameter of computing machine is Intel (R) Core (TM) i3CPU M350@2.27GHZ, internal memory 1.92GB.Video processing speed can reach real-time substantially, and for the more complicated video scene used in embodiment, the image time used of average treatment one frame 768 × 576 yardstick is about 60ms.
Above specific embodiments of the invention are described.It is to be appreciated that the present invention is not limited to above-mentioned particular implementation, those skilled in the art can make various distortion or amendment within the scope of the claims, and this does not affect flesh and blood of the present invention.
Claims (6)
1. detect a method for pedestrian target in video, it is characterized in that, the method comprises the steps:
The first step, gauss hybrid models is adopted to carry out modeling to the background of scene, by obtaining rough sport foreground with the comparison when pre-treatment image after obtaining background model, by sport foreground being proposed with the form of rectangular block after filtering and Morphological scale-space, be referred to as foreground blocks;
Second step, calculates foreground point proportion in each foreground blocks and distinguishes complicated foreground blocks and simple prospect block in conjunction with foreground point statistical property of projection number in horizontal coordinate, and estimating the horizontal coordinate that there is pedestrian;
3rd step, determines whether pedestrian for simple prospect block by its size; For complicated foreground blocks, utilize online pedestrian level estimation model and based on the Adaboost sorter of histograms of oriented gradients feature, pedestrian target accurately detected, obtaining the pedestrian target in video;
Described pedestrian's horizontal coordinate method of estimation is as follows: the foreground point utilizing gauss hybrid models to obtain projects in horizontal coordinate, the foreground point number that statistics drops in each horizontal coordinate generates histogram, to the smoothing filtering of the histogram obtained, the position that there is convex closure after filtering in histogram is exactly the position of pedestrian, estimates by carrying out convex closure detection to histogram the horizontal level that there is pedestrian;
Described simple prospect block refers to the foreground blocks comprising single body, complicated foreground blocks refers to the foreground blocks comprising multiple object, the mode distinguishing simple prospect block and complicated foreground blocks is as follows: for simple prospect block, shared by its foreground point pixel, the ratio of foreground blocks is more than or equal to 0.45, whether be simple prospect block by the threshold value preliminary judgement handling object of foregrounding point ratio, if it is determined that be simple prospect block, the method that recycling pedestrian horizontal coordinate is estimated checks wherein whether have more than one pedestrian target, wherein comprise multiple pedestrian target if estimated, foreground blocks is grouped in complicated foreground blocks, otherwise be simple prospect block, distinguishing a principle that is simple and complicated foreground blocks is: allow simple prospect block to be divided in complicated foreground blocks, but do not allow complicated foreground blocks to incorporate in simple prospect block,
Described online pedestrian level estimation model obtains in the following manner: assuming that the height Gaussian distributed of different pedestrians in scene, for its height in scene of same pedestrian with the change of vertical coordinate linear change, thus with a conditional Gaussian function p (h|y about vertical coordinate y, pedestrian level in scene β) is described, this model is not used to remove to estimate the height of pedestrian in the early stage, training Height Estimation model is removed as training sample with the height of the pedestrian target being judged to obtain by simple prospect block and coordinate information, p (h|y is recycled after being no less than the training of 30 times, β) pedestrian level of diverse location in scene is estimated, pedestrian level estimation model is
Wherein h (y) is the height of pedestrian target when ordinate is y, obeys the Gaussian distribution about condition y, and β is the parameter to be asked of model, σ
2for the distribution variance of specifying.
2. one kind adopts the detection method of particular color pedestrian under the Still Camera scene of method described in claim 1, it is characterized in that, described method in two steps, first adopts the first step described in claim 1 to detect the pedestrian target in video to the 3rd step, then identifies its clothing color.
3. the detection method of particular color pedestrian under Still Camera scene according to claim 2, it is characterized in that, described to its clothing color identify, specifically: after obtaining pedestrian target, utilize half body parted pattern by the upper body of pedestrian and the lower part of the body separated; Utilize method identification pedestrian's jacket of HSV space statistical color histogram and the color of trousers, finally mark the pedestrian target meeting setpoint color.
4. the detection method of particular color pedestrian under Still Camera scene according to claim 3, it is characterized in that, gradient map is calculated in described half body parted pattern, specifically: for the sport foreground point of pedestrian, in RGB tri-passages, calculate the gradient of vertical direction respectively according to Soble operator, after summation, obtain total gradient map; Obtain the Grad cumulative sum of often row pixel accordingly further, obtain the transverse projection histogram of gradient, then to this histogram analysis, just can obtain lower part of the body cut-off rule.
5. the detection method of particular color pedestrian under Still Camera scene according to claim 4, it is characterized in that, described cut-off rule drops in centre 1/3 scope of foreground blocks vertical direction.
6. the detection method of particular color pedestrian under Still Camera scene according to claim 3, it is characterized in that, described HSV space statistical color histogram method comprises HSV space segmentation and statistical color histogram two parts, the former refer to HSV space is divided into red, yellow, green, blue, purple, black, in vain, ash 8 color regions, the latter refers to the color conversion of carrying out from rgb space to HSV space to image, and statistics drops on the pixel number of above-mentioned 8 color regions, and then obtain the histogram reflecting color distribution, analyze histogram and draw clothing color.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210586125.8A CN103034852B (en) | 2012-12-28 | 2012-12-28 | The detection method of particular color pedestrian under Still Camera scene |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210586125.8A CN103034852B (en) | 2012-12-28 | 2012-12-28 | The detection method of particular color pedestrian under Still Camera scene |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103034852A CN103034852A (en) | 2013-04-10 |
| CN103034852B true CN103034852B (en) | 2015-10-28 |
Family
ID=48021731
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210586125.8A Expired - Fee Related CN103034852B (en) | 2012-12-28 | 2012-12-28 | The detection method of particular color pedestrian under Still Camera scene |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103034852B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103310444B (en) * | 2013-05-21 | 2016-02-10 | 上海交通大学 | A kind of method of the monitoring people counting based on overhead camera head |
| CN103530879B (en) * | 2013-10-15 | 2016-08-17 | 无锡清华信息科学与技术国家实验室物联网技术中心 | Pedestrian's color extraction method under special scenes |
| KR101498114B1 (en) * | 2013-11-28 | 2015-03-05 | 현대모비스 주식회사 | Device and method for detecting pedestrains |
| CN106844394B (en) * | 2015-12-07 | 2021-09-10 | 北京航天长峰科技工业集团有限公司 | Video retrieval method based on pedestrian clothes and shirt color discrimination |
| CN106056631A (en) * | 2016-06-06 | 2016-10-26 | 中国矿业大学 | Pedestrian detection method based on motion region |
| CN106407953A (en) * | 2016-10-14 | 2017-02-15 | 蔡璟 | Intelligent people searching system based on multi-image big data identification |
| CN106571014A (en) * | 2016-10-24 | 2017-04-19 | 上海伟赛智能科技有限公司 | Method for identifying abnormal motion in video and system thereof |
| CN106846329A (en) * | 2016-12-21 | 2017-06-13 | 天津中科智能识别产业技术研究院有限公司 | A kind of pedestrian's localization method for merging foreground detection |
| CN107145894B (en) * | 2017-03-13 | 2020-03-13 | 中山大学 | Target detection method based on directional gradient feature learning |
| CN108111783B (en) * | 2017-12-22 | 2020-08-14 | 河南英富迪光电科技有限公司 | Outdoor self-charging image recognition infrared imager |
| CN110298893A (en) * | 2018-05-14 | 2019-10-01 | 桂林远望智能通信科技有限公司 | A kind of pedestrian wears the generation method and device of color identification model clothes |
| CN115257625A (en) * | 2021-04-29 | 2022-11-01 | 上海博泰悦臻网络技术服务有限公司 | Safety belt control method and device and safety belt |
| CN114401365B (en) * | 2021-12-31 | 2024-05-14 | 广东省教育研究院 | Target person identification method, video switching method and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
| CN102799894A (en) * | 2012-06-21 | 2012-11-28 | 上海理工大学 | Detection method for rapidly overlooking pedestrians |
-
2012
- 2012-12-28 CN CN201210586125.8A patent/CN103034852B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
| CN102799894A (en) * | 2012-06-21 | 2012-11-28 | 上海理工大学 | Detection method for rapidly overlooking pedestrians |
Non-Patent Citations (1)
| Title |
|---|
| Robust Object Tracking by Combining Boosting Learning and Particle Filter;Tianyu LI etc.;《Journal of Computational Information Systems》;20121115;9593-9601 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103034852A (en) | 2013-04-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103034852B (en) | The detection method of particular color pedestrian under Still Camera scene | |
| CN104268583B (en) | Pedestrian re-recognition method and system based on color area features | |
| CN103310194B (en) | Pedestrian based on crown pixel gradient direction in a video shoulder detection method | |
| CN104050471B (en) | Natural scene character detection method and system | |
| CN102194108B (en) | Smile face expression recognition method based on clustering linear discriminant analysis of feature selection | |
| CN103310444B (en) | A kind of method of the monitoring people counting based on overhead camera head | |
| Sobottka et al. | Looking for faces and facial features in color images | |
| CN102902986A (en) | Automatic gender identification system and method | |
| CN105023008A (en) | Visual saliency and multiple characteristics-based pedestrian re-recognition method | |
| CN104951784A (en) | Method of detecting absence and coverage of license plate in real time | |
| CN108921004A (en) | Safety cap wears recognition methods, electronic equipment, storage medium and system | |
| CN103400110A (en) | Abnormal face detection method in front of ATM (automatic teller machine) | |
| CN103632132A (en) | Face detection and recognition method based on skin color segmentation and template matching | |
| CN108985170A (en) | Transmission line of electricity hanger recognition methods based on Three image difference and deep learning | |
| CN106127812B (en) | A kind of passenger flow statistical method of the non-gate area in passenger station based on video monitoring | |
| CN111127499A (en) | Security inspection image cutter detection segmentation method based on semantic contour information | |
| CN105205489A (en) | License plate detection method based on color texture analyzer and machine learning | |
| CN109460704A (en) | A kind of fatigue detection method based on deep learning, system and computer equipment | |
| CN103020614B (en) | Based on the human motion identification method that space-time interest points detects | |
| Galsgaard et al. | Circular hough transform and local circularity measure for weight estimation of a graph-cut based wood stack measurement | |
| CN104484645A (en) | Human-computer interaction-oriented '1' gesture-recognition method and system | |
| CN111553214A (en) | Method and system for detecting smoking behavior of driver | |
| Vishwakarma et al. | Simple and intelligent system to recognize the expression of speech-disabled person | |
| Kwaśniewska et al. | Face detection in image sequences using a portable thermal camera | |
| CN114581658A (en) | Target detection method and device based on computer vision |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20151028 Termination date: 20181228 |