CN103324955A - Pedestrian detection method based on video processing - Google Patents
Pedestrian detection method based on video processing Download PDFInfo
- Publication number
- CN103324955A CN103324955A CN2013102381272A CN201310238127A CN103324955A CN 103324955 A CN103324955 A CN 103324955A CN 2013102381272 A CN2013102381272 A CN 2013102381272A CN 201310238127 A CN201310238127 A CN 201310238127A CN 103324955 A CN103324955 A CN 103324955A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- image
- svm
- detection method
- foreground
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a pedestrian detection method based on video processing. The pedestrian detection method comprises the steps of (1) extracting a foreground image, extracting a moving object image of each frame of a video, marking the image and storing the image into a storage in sequence, using a background model to extract a background, enabling the model to adopt the gauss mixing model, (2) conducting preliminary screening on the foreground, selecting shape features of a pedestrian for conducting identification, (3) accurately identifying the foreground, selecting HOGs to conduct feature extraction on the foreground image after preliminary screening, then using a low dimensionality soft output SVM pedestrian classifier to conduct classification, and judging whether the pedestrian exists or not. The pedestrian detection method further comprises the step of (4) conducting error correction processing in a secondary thread. As for the foreground image with low dimensionality soft output SVM pedestrian classifier soft output results which are ambiguous in belonging classification, a high dimensionality SVM classifier is called in the secondary thread for recognition processing. The pedestrian detection method based on video processing improves the detection accuracy and is good in real-time performance.
Description
Technical field
The present invention relates to a kind of pedestrian detection method.
Background technology
Pedestrian detection is an important topic in the fields such as intelligent video monitoring, intelligent transportation, network picture/video frequency searching, can be applied directly in the application such as people counting, Pedestrian monitoring and warning, intelligent transportation pedestrian running red light, driver assistance safety driving system.The pedestrian detection algorithm is divided into two classes substantially: based on the pedestrian detection of still image with based on the pedestrian detection of Video processing.The former relies on pattern-recognition correlation technique identification pedestrian, and outstanding feature is that the accuracy of detection high real-time is poor.In present method, HOG feature and svm classifier device are the perfect adaptations of balance real-time and validity.The latter is partitioned into foreground target by pedestrian's movable information, further combined with the pedestrian the most intuitively shape facility (such as the ratio of width to height, area) identify, can remove simply, fast vehicle etc. is not pedestrian's target obviously, improves system effectiveness.But often degree of accuracy is very low in this recognition methods of not using any abstract characteristics of pedestrian.Therefore, pattern-recognition thought in the still image is incorporated in the Video processing, the degree of accuracy that increases pedestrian detection seems very urgent.In the recent period, also there are some scholars to do related work.Liu is more than the HOG feature extraction that has proposed to carry out respectively pyramid in " based on the video line people detection algorithm research of HOG feature and the movable information " literary composition that was published on " the Chinese science and technology paper is online " in 2009 under different scale.Although the variable size piece HOG characteristic block template that this algorithm adopts has increased the feature quantity of HOG testing result is increased, training and detection speed are slower.Yan Qing has introduced some how much invariant features such as the girth Area Ratio etc. in the conventional shape feature base in Shanghai Communications University's master thesis " vehicle in the monitor video and pedestrian detection " literary composition of delivering in 2009, and then re-uses sorter and accurately detect.The method improves detection speed by the prospect quantity that reduces input svm classifier device, does not fundamentally carry out algorithm and improves.
Summary of the invention
In order to overcome the deficiency that precision is lower, real-time is relatively poor of the detection that has pedestrian detection technology, the invention provides a kind of accuracy of detection, good pedestrian detection method based on Video processing of real-time of improving.
The technical solution adopted for the present invention to solve the technical problems is:
A kind of pedestrian detection method based on Video processing, described pedestrian detection method may further comprise the steps:
(1) extract foreground picture:
Extract the movement destination image of every frame in the video, and with its in order mark store container into; Use background modeling to extract background, described model adopts gauss hybrid models;
(2) preliminary screening prospect:
If target boundary rectangle area is δ, every frame movement destination image is of a size of W * H, and image area is S=W * H, and the video that obtains is divided into [0, x by horizontal or along slope coordinate
0), [x
0, x
1), [x
1, x
2] Three regions, judge whether δ/S satisfies threshold condition, if satisfy, be judged as for the first time the pedestrian;
Whether the ratio of width to height scale that again judges the target boundary rectangle meets the following conditions:
a<scale<b (5)
Wherein, a, b are the decimal between 0～1;
As satisfying, again be judged to be the pedestrian;
(3) accurately identify prospect: whether the foreground image after the preliminary screening is chosen HOG carry out feature extraction, then classify with the soft output of low dimension SVM pedestrian sorter, differentiating is the pedestrian.
Further, described pedestrian detection method is further comprising the steps of: correction process in (4) worker thread: the soft Output rusults of low dimension soft output SVM pedestrian's sorter is subordinate to the low foreground image of classification, calls high-dimensional svm classifier device and carry out again identifying processing in worker thread.
Further again, in the described step (4), extract low dimension HOG feature, the low soft output of the dimension SVM pedestrian sorter of input obtains for the first time pedestrian detection result, simultaneously, if the classification degree of membership of the soft output of low dimension SVM pedestrian sorter is low, storage foreground picture and this frame picture, then, enter the judgement of next frame picture.
In the described step (4), { 1,1} is mapped to [0,1] to the hard judgement Output rusults of the soft output of low dimension SVM pedestrian sorter, obtains being under the jurisdiction of pedestrian's probable value, probable value is belonged to the sample of setting threshold scope as the low sample of classification degree of membership.Described setting threshold scope can be [0.4,0.6], also can be other scopes.
Further, in the described step (4), when the quantity of storage foreground picture exceeds the present count value, triggering high-dimensional SVM pedestrian's sorter in the worker thread, whether be pedestrian, obtain for the second time result if differentiating; If the result is consistent with the soft output of low dimension SVM pedestrian sorter for the second time, then directly enter the judgement of next foreground image of storing, if result and the soft output of low dimension SVM pedestrian sorter are inconsistent for the second time, selecting for the second time, the result is final judged result.
Further again, in the described step (1), Gauss model is the feature that Gaussian probability-density function comes each pixel in the token image: establish G
tBe t background image constantly, each pixel on this moment background image set up mixed Gauss model:
Wherein, ω
M, t, μ
M, t,
Be respectively corresponding weight coefficient, average and variance,
Expression t is the distribution function of m gaussian component constantly;
Along with the variation of time, background image also can occur to change slowly, and mixed Gauss model needs to constantly update:
ω
m,t+1=(1+α)ω
m,t，μ
m,t+1=βμ
m,t+(1-β)I(x，y,t)，
(2)
Wherein, α is background model weight undated parameter, and β is background model average undated parameter, and I (x, y, t) is the gray-scale value that image I is located at pixel (x, y);
To mixed Gauss model, at moment t, each pixel attribute of formula (1) background model includes only two parameter: average μ and variances sigma.For a given image I (x, y, t), with each pixel I (x, y) and the corresponding constantly mixed Gauss model P (G of t
t) coupling, if satisfy
Wherein T chooses the interval Arbitrary Digit in 0.7≤T≤0.75 usually for the threshold value that we set, and thinks that then the match is successful, and (x, y) is judged as background dot; Otherwise (x, y) is the foreground point.
In the described step (2), described threshold condition is:
Beneficial effect of the present invention is mainly manifested in: improve accuracy of detection, real-time is good.
Description of drawings
Fig. 1 is the pedestrian detection method process flow diagram.
Fig. 2 is the schematic diagram of extraction prospect, wherein, (a) is former figure, (b) for mixed Gaussian extracts background, (c) is extraction prospect (165 frame), (d) is extraction prospect (210 frame).
Fig. 3 is prospect boundary rectangle schematic diagram, wherein, (a) is foreground target boundary rectangle (165 frame), (d) is foreground target boundary rectangle (210 frame).
Fig. 4 is pedestrian's recognition result schematic diagram, wherein, (a) is pedestrian's recognition result (165 frame), (d) is pedestrian's recognition result (210 frame).
Embodiment
The invention will be further described below in conjunction with accompanying drawing.
With reference to Fig. 1～Fig. 4, a kind of pedestrian detection method based on Video processing, described pedestrian detection method may further comprise the steps:
(1) extract foreground picture:
Extract the movement destination image of every frame in the video, and with its in order mark store container into; Use background modeling to extract background, described model adopts gauss hybrid models;
(2) preliminary screening prospect:
If target boundary rectangle area is δ, every frame movement destination image is of a size of W * H, and image area is S=W * H, and the video that obtains is divided into [0, x by horizontal or along slope coordinate
0), [x
0, x
1), [x
1, x
2] Three regions, judge whether δ/S satisfies threshold condition, if satisfy, be judged as for the first time the pedestrian;
Whether the ratio of width to height scale that again judges the target boundary rectangle meets the following conditions:
a<scale<b (5)
Wherein, a, b are the decimal between 0～1;
As satisfying, again be judged to be the pedestrian;
(3) accurately identify prospect: whether the foreground image after the preliminary screening is chosen HOG carry out feature extraction, then classify with the soft output of low dimension SVM pedestrian sorter, differentiating is the pedestrian.
Further, described pedestrian detection method is further comprising the steps of: correction process in (4) worker thread: the low soft Output rusults of the soft output of dimension SVM pedestrian's sorter is subordinate to the indefinite foreground image of classification, calls high-dimensional svm classifier device and carry out again identifying processing in worker thread.
The pedestrian detecting system of the present embodiment is comprised of main thread module and worker thread module two parts.Main thread functions of modules: at first, realize loading, displaying video or be attached thereto the camera that connects; Then, poor obtain the foreground target in the video of loading by doing with background, and utilize pedestrian's shape facility (such as the ratio of width to height, area) preliminary screening to go out the pedestrian in the target; Then utilize and be input as low dimension HOG feature, be output as SVM pedestrian's sorter of the probable value that is under the jurisdiction of the pedestrian to further accurately identification of prospect, simultaneously, all are subordinate to the indefinite foreground target sequential storage of classification in formation; At last, the result of output main thread module pedestrian detection.
Worker thread functions of modules: call the higher svm classifier device of input feature vector dimension indefinite foreground target of classifying being subordinate to of main thread storage is carried out again identifying processing.Consider that the computing machine thread utilizes efficient problem, we only have just trigger one time worker thread when storage foreground target number reach 10, after utilizing high-dimensional SVM to handle 10 pictures and carrying out identifying processing again, worker thread is closed, and waits for triggering next time.The below describes in detail to each module principle.
Extract foreground picture: the extraction prospect is namely extracted the moving target of every frame in the video, and with its in order mark store in the container.In the Video processing that camera is fixed, the detection method of moving target can be summarized as three classes generally: optical flow method, frame-to-frame differences method and background subtraction method.This paper adopts background subtraction point-score simple and that be easy to realize.
The key of background difference is to use background modeling to extract background.The method of background modeling is a lot, but the background after the general modeling is not very clean clear, and gauss hybrid models is one of the most successful method of modeling.It is the feature that Gaussian probability-density function comes each pixel in the token image with Gauss model: establish G
tBe t background image constantly, each pixel on this moment background image set up mixed Gauss model:
Wherein,
Be respectively corresponding weight coefficient, average and variance,
Expression t is the distribution function of m gaussian component constantly.
Along with the variation of time, background image also can occur to change slowly, and mixed Gauss model needs to constantly update:
ω
m,t+1=(1+α)ω
m,t,μ
m,t+1=βμ
m,t+(1-β)I(x,y,t)，
(2)
Wherein, α is background model weight undated parameter, and β is background model average undated parameter, and I (x, y, t) is the gray-scale value that image I is located at pixel (x, y).
To mixed Gauss model, at moment t, each pixel attribute of formula (1) background model includes only two parameter: average μ and variances sigma.For a given image I (x, y, t), with each pixel I (x, y) and the corresponding constantly mixed Gauss model P (G of t
t) coupling, if satisfy
Wherein T chooses the interval Arbitrary Digit in 0.7≤T≤0.75 usually for the threshold value that we set, and thinks that then the match is successful, and (x, y) is judged as background dot; Otherwise (x, y) is the foreground point.
The preliminary screening prospect: pedestrian's identification is by some feature of extracting the pedestrian a series of moving targets that image segmentation obtains to be identified, classified, and the target identification that is judged as the pedestrian out.The preliminary screening prospect be choose the pedestrian the most intuitively shape facility (such as the ratio of width to height, area) identify, can remove simply, fast vehicle etc. is not pedestrian's target obviously, improve system effectiveness, but this recognition methods of not using any abstract characteristics of pedestrian often degree of accuracy can be very low.
If target boundary rectangle area is δ, every two field picture is of a size of W * H, and image area is S=W * H.Consider that same motion is different from the camera distance, area also has marked difference, so we are divided into [0, x to the video that obtains by horizontal or along slope coordinate
0), [x
0, x
1), [x
1, x
2] Three regions.Judge whether δ/S satisfies
If satisfy (4), be judged as for the first time the pedestrian, cross general objective and the too small target such as leaf, antenna to remove the vehicle homalographic.Further, if the ratio of width to height scale of target boundary rectangle satisfies
Again be judged to be the pedestrian.Because the tall feature of pedestrian is fairly obvious, therefore can finely isolate the pedestrian with this ratio.
Accurate identification prospect: in order to improve the pedestrian detection accuracy rate, reject mistake, this paper introduces the low soft output of dimension SVM pedestrian sorter on the basis of classic method and implements accurate pedestrian detection.
SVM is a kind of based on statistical valid data sorting machine learning method.It is better than other sorters to the classification performance of two class problems.The HOG feature can well characterize local object appearance and shape, and to illumination-insensitive.The people such as Dalal successively adopt MIT and INRIA pedestrian's picture library, obtain the HOG proper vector of different pictures as positive negative sample, send into SVM and train, can be 10
-4Realize the verification and measurement ratio more than 80% under the fallout ratio.So HOG and SVM can be rated as the most perfect combination in the static images pedestrian detection algorithm.The soft output of low dimension in this paper pedestrian sorter is chosen HOG and is carried out feature extraction, then classifies with SVM.Training classifier and be presented below with the principle of the sorter that trains identification prospect.
The low soft output of the dimension SVM pedestrian sorter principle of training:
1) chooses sample set
Pedestrian's image data base commonly used has MIT database, Benchmark database, each ITS project team voluntarily acquisition database and INRIA database.Wherein the INRIA database has comprised colored personage's photo of 970MB, comprises various attitudes, such as climbing the mountain, squat down etc.Also have in addition a large amount of complete background pictures without pedestrian detection, can be used for the extraction of magnanimity negative sample.Therefore, we can select to be normalized to 64 * 128 positive and negative sample set and carry out the training of sorter from the INRIA database.
2) according to foreground picture feature and sorter characteristic, carry out the normalized of suitable dimension.The foreground target picture that extracts has following characteristics with respect to static images arbitrarily: 1) dimension of picture is little, is zonule in every frame picture; 2) the target proportion is larger in the picture, and namely characteristic information is more.Because dimension of picture is little, the intrinsic dimensionality of extraction can reduce greatly, and the sorting algorithm complexity reduces thereupon.To any static images, the zone outside the target not only aligns the classification of negative sample without any help, can produce the classification of sorter on the contrary and disturb.So, the picture of same size, main information is more, and the sorter Output rusults is more reliable.
Based on this, we do following pre-service to the sample set of choosing: align sample, contain pedestrian's zone in the cut-away view, and normalize to 32 * 64 sizes; Negative sample directly is normalized to 32 * 64.
3) extract sample set HOG feature, the low soft output of the dimension SVM pedestrian sorter of training.
(a) to the sample piecemeal, block size is 16 * 16 pixels; The unit that each piece is divided into 48 * 8 pixels, step-length are 8 pixels.So after the normalized, each piece be characterized as 36 dimensional vectors that the feature cascade with 4 unit obtains.Every pictures HOG intrinsic dimensionality before the dimensionality reduction: have 7 * 15 pieces, therefore total dimension is 7 * 15 * 36=3780 dimension.Every pictures HOG intrinsic dimensionality behind the dimensionality reduction: have 3 * 7 pieces, therefore total dimension is 3 * 7 * 36=756 dimension.To sum up, computational complexity has reduced by 5 times.
(b) adopt the LIBSVM sorter, and { 1,1} is mapped to [0,1] the hard judgement Output rusults of LIBSVM.
LIBSVM is by one of Taiwan Univ. exploitation svm classifier device fast and effectively.The file that has compiled and can directly carry out in the Windows system is not only arranged, source code also is provided, the convenient improvement.SVM is input as the sample characteristics of extraction, be output as classification judge information 1,1}, general " 1 " expression does not belong to this classification, " showing " expression belongs to this classification.But because the pollution of noise etc. can't be classified as a certain class to the sample that is in the class border clearly, can only differentiate as a certain class take certain probability or certain degree of membership during practical application.Therefore, in order to make the svm classifier device more be applicable to these problems, the t J.C. of the publication that Platt J.C. published in 1999, time spent is because the pollution of noise etc. can't be classified as the concept that has proposed soft output (soft output) in the book in " Probabilistic Output for Support Vector Machine and Comparisons to Regularized Likelihood Methods " literary composition to the sample that is in the class border clearly.As continuous function the hard judgement output f (x) of SVM is mapped to [0,1] with an extruding function sigmoid (), realizes the posterior probability output of SVM.
We introduce this thought in low dimension SVM pedestrian sorter, judge that hard { 1,1} is mapped to [0,1] to Output rusults, is called the probable value that is under the jurisdiction of the pedestrian.And the sample separation that probable value is belonged to [0.4,0.6] out, uses multithreading thought, calls a high-dimensional svm classifier device and further carry out pedestrian's identification in another thread, draws accurate judged result.Accurate pedestrian's identification is analyzed necessary for follow-up pedestrian behavior.
(c) the samples pictures HOG feature of extracting in (a) is input in the amended LIBSVM sorter, trains, obtain pedestrian's sorter of the low soft output of dimension.
The sorter identification foreground picture that use trains:
1) with the low soft output of dimension SVM pedestrian sorter identification pedestrian.As a width of cloth independent picture, through normalization size HOG feature extraction, whether the low soft output svm classifier of the dimension device of input successively differentiates the pedestrian each moving target.The value of output reliability parameter.
2) separate the low picture of reliability and storage.
The sample that dependability parameter belongs to [0.4,0.6] is separated, and stores isolated Target Photo, and records the moment of this picture place frame.
Correction process in the worker thread: the purpose of setting up worker thread is in order further to improve detection accuracy not affecting to detect on the real-time basis, in analyzing to the pedestrian behavior that better is applied to the back.What call in the worker thread is the SVM pedestrian's sorter that obtains with the higher HOG features training of precision.The normalization that it is chosen is of a size of 64 * 128, and the HOG that each foreground image extracts is characterized as 3780 dimensions.So, if in main thread, directly call this sorter, will certainly exert an influence to the algorithm real-time.And worker thread can normally move under the prerequisite not affecting main thread, helps it to finish pedestrian detection result's again identifying processing.
In the master routine processing procedure, soft Output rusults is subordinate to the corresponding of the indefinite foreground target of classification and place frame picture and has constantly stored in the formation that defines.Worker thread is processed these pictures, and principle of work is as follows:
When 1) quantity of queue stores foreground target is greater than 10, triggers worker thread and start working.
2) read and the normalization formation in foreground picture.From formation, read the foreground picture of storage by the order of first in first out, and normalization is of a size of 64 * 128.
3) extraction foreground picture HOG feature is input in the svm classifier device that trains and detects the output detections result.
Compare with the soft output of low dimension SVM pedestrian sorter Output rusults, if consistent, the judgement that directly enters next pictures, inconsistent, adopt new result to replace it front result.
The program running environment of the present embodiment is under Visual Studio2010 development platform, a MFC interface based on dialog box of establishment.In the algorithm implementation procedure, called image processing related function among the OpenCV2.3.1.
The treatment scheme of the pedestrian detection method of the present embodiment is as follows:
1) loads, closes the pedestrian detection video.Video is image one by one, according to the frame per second of video suitable clock is set, and regularly the two field picture in the video is carried out reading displayed.Loading video can realize by functions such as DrawToHDC () among the OpenCV; Closing video can realize by a label, a label namely is set is named as m_run, and when its value is 0, picture zero setting in the viewing area.
2) extraction prospect.Be to carry out background modeling with mixed Gauss model in the literary composition, do poor and carry out the operations such as thresholding, morphologic filtering and obtain the two-value foreground picture with present frame and background.Can be by calling morphologyEx () among the OpenCV, erode (), the realization of dilate () function.
3) rim detection and obtain boundary rectangle.Rim detection is to obtain the profile coordinate of every frame foreground target, and then obtains each profile boundary rectangle coordinate, is stored as vector＜Rect〉type.Can be by calling findContours () among the OpenCV, the realization of boundingRect () function.The pedestrian of back identification all is to vector＜Rect〉boundary rectangle of type operates.
4) preliminary screening prospect.Adopt pedestrian's shape information to screen.Calculate each target area δ, satisfy formula (4), be judged as for the first time the pedestrian, cross general objective and the too small target such as leaf, antenna to remove the vehicle homalographic.The ratio of width to height scale that calculates the target boundary rectangle satisfies formula (5), again is judged as the pedestrian.
5) accurately identify prospect.
A) call resize () function the input foreground picture is normalized to 32 * 64.
B) extract foreground picture HOG feature, dimension is 756.
C) be input in pedestrian's sorter of the soft output of low dimension that has trained and detect output pedestrian detection result.
Be input in pedestrian's sorter of the soft output of low dimension that trains.Be input to again in the pedestrian's sorter that has trained.At last, mark out the pedestrian who identifies with rectangle ().
6) separation is subordinate to the indefinite foreground target of classification and storage.The storage probable value in the foreground picture of [0.4,0.6] in the formation that defines, and moment of recording this picture place frame.
7) correction process in the worker thread.
When a) quantity of storage foreground target is greater than 10, triggers second thread and start working.
B) from formation, read the foreground picture of storage by the order of first in first out, call resize () function normalization size to 64 * 128, the HOG feature of extracting 3780 dimensions is input to high-dimensional svm classifier device, export more reliable pedestrian's recognition result, compare with the soft output of low dimension SVM pedestrian sorter Output rusults.If consistent, the judgement that directly enters next pictures; If inconsistent, cover event memory before with existing Output rusults.
8) display line people detection result.The boundary rectangle of the foreground target that is judged as the pedestrian can be drawn, rectangle () function realization among the OpenCV can be called.
In addition, before service routine is carried out pedestrian detection first, select first the training of suitable sample set to obtain required sorter, the low dimension that arrives used herein soft output pedestrian's svm classifier device and high-dimensional svm classifier device training step are as follows.
The low soft output of the dimension pedestrian sorter step of training:
1) from the INRIA sample set, chooses and be of a size of that 64 * 128 800 width of cloth do not repeat positive sample and 500 width of cloth do not repeat negative sample, be used for training pedestrian sorter.
2) align sample, contain pedestrian's zone in the cut-away view, and normalize to 32 * 64; Negative sample directly is normalized to 32 * 64.
3) the HOG feature of extraction samples pictures.Block size is 16 * 16 pixels; Each piece is equally divided into 4 unit, and step-length is 8 pixels.The HOG intrinsic dimensionality of every pictures is 756 dimensions.
4) obtain LIBSVM sorter source code.Downloading network address is http://www.csie.ntu.edu.tw/～cjlin/, and finishes on request data and prepare and application configuration work.
5) change parameter, { 1,1} is mapped to the probable value of [0,1] the hard judgement Output rusults of LIBSVM.In LIBSVM, by the training of svmtrain realization to training dataset, obtain the SVM model.Use-pattern is: svmtrain[options] training_set_file[model_file], wherein, options is operating parameter, the modification of the output type of SVM, kernel function type etc. all is to arrange herein.
6) the samples pictures HOG feature of extracting is input in the amended LIBSVM sorter, trains, obtain pedestrian's sorter of the low soft output of dimension.
High-dimensional svm classifier device in the training worker thread: step is with the training of general svm classifier device, and the normalization that we choose at this is of a size of the most frequently used 64 * 128.
Be each Implement of Function Module effect of checking this paper algorithm, we choose one section garden video that the tiny motion artifacts such as car, people, rope yarn are arranged, video size 480 * 360,25 frame/seconds of frame per second.Program is moved at PC, and the PC configuration: central processing unit (CPU) is Intel i3 platform; Internal memory (Memory) is 2GB DDR3.The detection effect analysis is as follows.
Fig. 2 is the background picture and different constantly foreground picture that uses mixed Gauss model to extract.Background picture is more complete among Fig. 2 b, and just the right side part of rope yarn is differentiated for prospect because shake.Fig. 2 c, the prospect that extracts among the 2d has all moving targets such as car, pedestrian, part rope yarn, and wherein trickle cavity is the shortcoming that the background subtraction point-score is difficult to avoid.
Fig. 3 is the promising boundary rectangle that extracts in the different moment; Fig. 4 is the boundary rectangle through the foreground target after the again identification in pedestrian's area mentioned above, the ratio of width to height, the low soft output category device of dimension and the worker thread.Find through contrast, pedestrian's recognition methods of this paper can be screened non-pedestrian's foreground targets such as vehicle among Fig. 3, and then the complete pedestrian zone that detects, and can finely be applied to during subsequent pedestrian behavior analyzes.The method processing speed also than comparatively fast, can satisfy the practical application request of real-time.
Claims (6)
1. pedestrian detection method based on Video processing, it is characterized in that: described pedestrian detection method may further comprise the steps:
(1) extract foreground picture:
Extract the movement destination image of every frame in the video, and with its in order mark store container into; Use background modeling to extract background, described model adopts gauss hybrid models;
(2) preliminary screening prospect:
If target boundary rectangle area is δ, every frame movement destination image is of a size of W * H, and image area is S=W * H, and the video that obtains is divided into [0, x by horizontal or along slope coordinate
0), [x
0, x
1), [x
1, x
2] Three regions, judge whether δ/S satisfies threshold condition, if satisfy, be judged as for the first time the pedestrian;
Whether the ratio of width to height scale that again judges the target boundary rectangle meets the following conditions:
a<scale<b (5)
Wherein, a, b are the decimal between 0～1;
As satisfying, again be judged to be the pedestrian;
(3) accurately identify prospect: whether the foreground image after the preliminary screening is chosen HOG carry out feature extraction, then classify with the soft output of low dimension SVM pedestrian sorter, differentiating is the pedestrian.
2. a kind of pedestrian detection method based on Video processing as claimed in claim 1, it is characterized in that: described pedestrian detection method is further comprising the steps of: correction process in (4) worker thread: the soft Output rusults of low dimension soft output SVM pedestrian's sorter is subordinate to the low foreground image of classification, calls high-dimensional svm classifier device and carry out again identifying processing in worker thread.
3. a kind of pedestrian detection method based on Video processing as claimed in claim 2, it is characterized in that: in the described step (4), extract low dimension HOG feature, the low soft output of the dimension SVM pedestrian sorter of input obtains for the first time pedestrian detection result, simultaneously, if the classification degree of membership of the soft output of low dimension SVM pedestrian sorter is low, storage foreground picture and this frame picture, then, enter the judgement of next frame picture.
4. a kind of pedestrian detection method based on Video processing as claimed in claim 2 or claim 3, it is characterized in that: in the described step (4), when the quantity of storage foreground picture exceeds the present count value, trigger high-dimensional SVM pedestrian's sorter in the worker thread, whether be pedestrian, obtain for the second time result if differentiating; If the result is consistent with the soft output of low dimension SVM pedestrian sorter for the second time, then directly enter the judgement of next foreground image of storing, if result and the soft output of low dimension SVM pedestrian sorter are inconsistent for the second time, selecting for the second time, the result is final judged result.
5. a kind of pedestrian detection method based on Video processing as claimed in claim 1 or 2, it is characterized in that: in the described step (1), Gauss model is the feature that Gaussian probability-density function comes each pixel in the token image: establish G
tBe t background image constantly, each pixel on this moment background image set up mixed Gauss model:
Wherein, ω
M, t, μ
M, t,
Be respectively corresponding weight coefficient, average and variance,
Expression t is the distribution function of m gaussian component constantly;
Along with the variation of time, background image also can occur to change slowly, and mixed Gauss model needs to constantly update:
ω
m,t+1=(1+α)ω
m,t,μ
m,t+1=βμ
m,t+(1-β)I(x，y,t)，
(2)
Wherein, α is background model weight undated parameter, and β is background model average undated parameter, and I (x, y, t) is the gray-scale value that image I is located at pixel (x, y);
To mixed Gauss model, at moment t, each pixel attribute of formula (1) background model includes only two parameter: average μ and variances sigma, for a given image I (x, y, t), with each pixel I (x, y) and corresponding constantly rare mixed Gauss model P (G
t) coupling, if satisfy
Wherein T chooses the interval Arbitrary Digit in 0.7≤T≤0.75 usually for the threshold value that we set, and thinks that then the match is successful, and (x, y) is judged as background dot; Otherwise (x, y) is the foreground point.
6. a kind of pedestrian detection method based on Video processing as claimed in claim 1 or 2, it is characterized in that: in the described step (2), described threshold condition is:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013102381272A CN103324955A (en) | 2013-06-14 | 2013-06-14 | Pedestrian detection method based on video processing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013102381272A CN103324955A (en) | 2013-06-14 | 2013-06-14 | Pedestrian detection method based on video processing |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103324955A true CN103324955A (en) | 2013-09-25 |
Family
ID=49193683
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2013102381272A Pending CN103324955A (en) | 2013-06-14 | 2013-06-14 | Pedestrian detection method based on video processing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103324955A (en) |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103632143A (en) * | 2013-12-05 | 2014-03-12 | 冠捷显示科技（厦门）有限公司 | Cloud computing combined object identifying system on basis of images |
| CN104063722A (en) * | 2014-07-15 | 2014-09-24 | 国家电网公司 | Safety helmet identification method integrating HOG human body target detection and SVM classifier |
| CN104751485A (en) * | 2015-03-20 | 2015-07-01 | 安徽大学 | GPU adaptive foreground extracting method |
| CN105184818A (en) * | 2015-09-06 | 2015-12-23 | 山东华宇航天空间技术有限公司 | Video monitoring abnormal behavior detection method and detections system thereof |
| CN105224922A (en) * | 2015-09-18 | 2016-01-06 | 桂林远望智能通信科技有限公司 | A kind of people's car classification video streaming disposal system and method |
| CN105574515A (en) * | 2016-01-15 | 2016-05-11 | 南京邮电大学 | Pedestrian re-identification method in zero-lap vision field |
| WO2016106595A1 (en) * | 2014-12-30 | 2016-07-07 | Nokia Technologies Oy | Moving object detection in videos |
| WO2016111640A1 (en) * | 2015-01-07 | 2016-07-14 | Viscando Ab | Method and system for categorization of a scene |
| CN106210674A (en) * | 2016-09-22 | 2016-12-07 | 江苏理工学院 | Personnel-oriented monitoring video data file rapid processing method and system |
| CN106355162A (en) * | 2016-09-23 | 2017-01-25 | 江西洪都航空工业集团有限责任公司 | Method for detecting intrusion on basis of video monitoring |
| CN106407926A (en) * | 2016-09-09 | 2017-02-15 | 东软集团股份有限公司 | Pedestrian detection method and apparatus |
| CN106650577A (en) * | 2016-09-22 | 2017-05-10 | 江苏理工学院 | Method and system for quickly searching target person in monitoring video data file |
| CN107301380A (en) * | 2017-06-01 | 2017-10-27 | 华南理工大学 | One kind is used for pedestrian in video monitoring scene and knows method for distinguishing again |
| CN107734799A (en) * | 2017-10-19 | 2018-02-23 | 南京工程学院 | A kind of intelligent lighting system and its control method of accessible video monitoring |
| CN107872644A (en) * | 2016-09-23 | 2018-04-03 | 亿阳信通股份有限公司 | Video frequency monitoring method and device |
| CN108038872A (en) * | 2017-12-22 | 2018-05-15 | 中国海洋大学 | One kind perceives follow method based on sound state target detection and Real Time Compression |
| CN108229411A (en) * | 2018-01-15 | 2018-06-29 | 上海交通大学 | Human body hand-held knife behavioral value system and method based on RGB color image |
| CN108647649A (en) * | 2018-05-14 | 2018-10-12 | 中国科学技术大学 | The detection method of abnormal behaviour in a kind of video |
| CN109086681A (en) * | 2018-07-11 | 2018-12-25 | 东南大学 | A kind of intelligent video black smoke vehicle detection method based on LHI feature |
| CN109359549A (en) * | 2018-09-20 | 2019-02-19 | 广西师范大学 | A kind of pedestrian detection method based on mixed Gaussian and HOG_LBP |
| CN109684946A (en) * | 2018-12-10 | 2019-04-26 | 成都睿码科技有限责任公司 | A kind of kitchen mouse detection method based on the modeling of single Gaussian Background |
| CN109816004A (en) * | 2019-01-17 | 2019-05-28 | 平安城市建设科技(深圳)有限公司 | Source of houses picture classification method, device, equipment and storage medium |
| CN109840982A (en) * | 2019-01-02 | 2019-06-04 | 京东方科技集团股份有限公司 | It is lined up recommended method and device, computer readable storage medium |
| CN109948432A (en) * | 2019-01-29 | 2019-06-28 | 江苏裕兰信息科技有限公司 | A kind of pedestrian detection method |
| CN109977844A (en) * | 2019-03-21 | 2019-07-05 | 江苏科技大学 | One kind giving precedence to pedestrian's control system and control method based on image recognition motor vehicle |
| CN110472569A (en) * | 2019-08-14 | 2019-11-19 | 旭辉卓越健康信息科技有限公司 | A kind of method for parallel processing of personnel detection and identification based on video flowing |
| CN110599781A (en) * | 2019-09-23 | 2019-12-20 | 潍坊畅讯交通设施工程有限公司 | Traffic flow statistics and recognition system for intelligent city |
| CN110866421A (en) * | 2018-08-28 | 2020-03-06 | 天津理工大学 | Pedestrian identification method based on light field camera and HOG and SIFT mixed features |
| CN110866420A (en) * | 2018-08-28 | 2020-03-06 | 天津理工大学 | 2D deceptive pedestrian recognition method based on light field camera, HOG and SVM |
| CN113034551A (en) * | 2021-05-31 | 2021-06-25 | 南昌虚拟现实研究院股份有限公司 | Target tracking and labeling method and device, readable storage medium and computer equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120147189A1 (en) * | 2010-12-08 | 2012-06-14 | GM Global Technology Operations LLC | Adaptation for clear path detection using reliable local model updating |
| CN102147869B (en) * | 2011-03-31 | 2012-11-28 | 上海交通大学 | Pedestrian detection method based on foreground analysis and pattern recognition |
-
2013
- 2013-06-14 CN CN2013102381272A patent/CN103324955A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120147189A1 (en) * | 2010-12-08 | 2012-06-14 | GM Global Technology Operations LLC | Adaptation for clear path detection using reliable local model updating |
| CN102147869B (en) * | 2011-03-31 | 2012-11-28 | 上海交通大学 | Pedestrian detection method based on foreground analysis and pattern recognition |
Non-Patent Citations (4)
| Title |
|---|
| 刘晓敏: "基于神经网络和D-S证据理论的交通目标识别", 《计算机仿真》, vol. 27, no. 9, 30 September 2010 (2010-09-30), pages 295 - 298 * |
| 李娟等: "基于混合高斯模型的行人检测方法", 《吉林大学学报（工学版）》, vol. 41, no. 1, 31 January 2011 (2011-01-31), pages 41 - 45 * |
| 谢立等: "面向智能视频监控系统运动目标检测的轮廓提取方法", 《东南大学学报(自然科学版)》, vol. 42, 30 September 2012 (2012-09-30), pages 31 - 35 * |
| 贾俊涛: "一种串联组合分类的红外行人检测方法", 《应用光学》, vol. 33, no. 4, 31 July 2012 (2012-07-31), pages 770 - 773 * |
Cited By (41)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103632143B (en) * | 2013-12-05 | 2017-02-08 | 冠捷显示科技（厦门）有限公司 | Cloud computing combined object identifying system on basis of images |
| CN103632143A (en) * | 2013-12-05 | 2014-03-12 | 冠捷显示科技（厦门）有限公司 | Cloud computing combined object identifying system on basis of images |
| CN104063722A (en) * | 2014-07-15 | 2014-09-24 | 国家电网公司 | Safety helmet identification method integrating HOG human body target detection and SVM classifier |
| CN104063722B (en) * | 2014-07-15 | 2018-03-23 | 国家电网公司 | A kind of detection of fusion HOG human body targets and the safety cap recognition methods of SVM classifier |
| WO2016106595A1 (en) * | 2014-12-30 | 2016-07-07 | Nokia Technologies Oy | Moving object detection in videos |
| CN107209941A (en) * | 2014-12-30 | 2017-09-26 | 诺基亚技术有限公司 | Mobile object detection in video |
| US10127458B2 (en) | 2015-01-07 | 2018-11-13 | Viscando Ab | Method and system for categorization of a scene |
| WO2016111640A1 (en) * | 2015-01-07 | 2016-07-14 | Viscando Ab | Method and system for categorization of a scene |
| CN104751485A (en) * | 2015-03-20 | 2015-07-01 | 安徽大学 | GPU adaptive foreground extracting method |
| CN104751485B (en) * | 2015-03-20 | 2017-05-24 | 安徽大学 | GPU adaptive foreground extracting method |
| CN105184818B (en) * | 2015-09-06 | 2018-05-18 | 山东华宇航天空间技术有限公司 | A kind of video monitoring anomaly detection method and its detecting system |
| CN105184818A (en) * | 2015-09-06 | 2015-12-23 | 山东华宇航天空间技术有限公司 | Video monitoring abnormal behavior detection method and detections system thereof |
| CN105224922B (en) * | 2015-09-18 | 2018-10-16 | 桂林远望智能通信科技有限公司 | A kind of people's vehicle classification video streaming processing system and method |
| CN105224922A (en) * | 2015-09-18 | 2016-01-06 | 桂林远望智能通信科技有限公司 | A kind of people's car classification video streaming disposal system and method |
| CN105574515B (en) * | 2016-01-15 | 2019-01-01 | 南京邮电大学 | A kind of pedestrian recognition methods again under non-overlapping visual field |
| CN105574515A (en) * | 2016-01-15 | 2016-05-11 | 南京邮电大学 | Pedestrian re-identification method in zero-lap vision field |
| CN106407926B (en) * | 2016-09-09 | 2019-09-20 | 东软集团股份有限公司 | A kind of pedestrian detection method and device |
| CN106407926A (en) * | 2016-09-09 | 2017-02-15 | 东软集团股份有限公司 | Pedestrian detection method and apparatus |
| CN106650577A (en) * | 2016-09-22 | 2017-05-10 | 江苏理工学院 | Method and system for quickly searching target person in monitoring video data file |
| CN106210674A (en) * | 2016-09-22 | 2016-12-07 | 江苏理工学院 | Personnel-oriented monitoring video data file rapid processing method and system |
| CN106355162A (en) * | 2016-09-23 | 2017-01-25 | 江西洪都航空工业集团有限责任公司 | Method for detecting intrusion on basis of video monitoring |
| CN107872644A (en) * | 2016-09-23 | 2018-04-03 | 亿阳信通股份有限公司 | Video frequency monitoring method and device |
| CN107301380A (en) * | 2017-06-01 | 2017-10-27 | 华南理工大学 | One kind is used for pedestrian in video monitoring scene and knows method for distinguishing again |
| CN107734799A (en) * | 2017-10-19 | 2018-02-23 | 南京工程学院 | A kind of intelligent lighting system and its control method of accessible video monitoring |
| CN108038872A (en) * | 2017-12-22 | 2018-05-15 | 中国海洋大学 | One kind perceives follow method based on sound state target detection and Real Time Compression |
| CN108038872B (en) * | 2017-12-22 | 2021-08-31 | 中国海洋大学 | Dynamic and static target detection and real-time compressed sensing tracking research method |
| CN108229411A (en) * | 2018-01-15 | 2018-06-29 | 上海交通大学 | Human body hand-held knife behavioral value system and method based on RGB color image |
| CN108647649A (en) * | 2018-05-14 | 2018-10-12 | 中国科学技术大学 | The detection method of abnormal behaviour in a kind of video |
| CN108647649B (en) * | 2018-05-14 | 2021-10-01 | 中国科学技术大学 | Method for detecting abnormal behaviors in video |
| CN109086681A (en) * | 2018-07-11 | 2018-12-25 | 东南大学 | A kind of intelligent video black smoke vehicle detection method based on LHI feature |
| CN110866421A (en) * | 2018-08-28 | 2020-03-06 | 天津理工大学 | Pedestrian identification method based on light field camera and HOG and SIFT mixed features |
| CN110866420A (en) * | 2018-08-28 | 2020-03-06 | 天津理工大学 | 2D deceptive pedestrian recognition method based on light field camera, HOG and SVM |
| CN109359549A (en) * | 2018-09-20 | 2019-02-19 | 广西师范大学 | A kind of pedestrian detection method based on mixed Gaussian and HOG_LBP |
| CN109684946A (en) * | 2018-12-10 | 2019-04-26 | 成都睿码科技有限责任公司 | A kind of kitchen mouse detection method based on the modeling of single Gaussian Background |
| CN109840982A (en) * | 2019-01-02 | 2019-06-04 | 京东方科技集团股份有限公司 | It is lined up recommended method and device, computer readable storage medium |
| CN109816004A (en) * | 2019-01-17 | 2019-05-28 | 平安城市建设科技(深圳)有限公司 | Source of houses picture classification method, device, equipment and storage medium |
| CN109948432A (en) * | 2019-01-29 | 2019-06-28 | 江苏裕兰信息科技有限公司 | A kind of pedestrian detection method |
| CN109977844A (en) * | 2019-03-21 | 2019-07-05 | 江苏科技大学 | One kind giving precedence to pedestrian's control system and control method based on image recognition motor vehicle |
| CN110472569A (en) * | 2019-08-14 | 2019-11-19 | 旭辉卓越健康信息科技有限公司 | A kind of method for parallel processing of personnel detection and identification based on video flowing |
| CN110599781A (en) * | 2019-09-23 | 2019-12-20 | 潍坊畅讯交通设施工程有限公司 | Traffic flow statistics and recognition system for intelligent city |
| CN113034551A (en) * | 2021-05-31 | 2021-06-25 | 南昌虚拟现实研究院股份有限公司 | Target tracking and labeling method and device, readable storage medium and computer equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103324955A (en) | Pedestrian detection method based on video processing | |
| US8824742B2 (en) | Occupancy detection for managed lane enforcement based on localization and classification of windshield images | |
| CN102163290B (en) | Method for modeling abnormal events in multi-visual angle video monitoring based on temporal-spatial correlation information | |
| US20160078306A1 (en) | System and method for detecting seat belt violations from front view vehicle images | |
| CN201255897Y (en) | Human flow monitoring device for bus | |
| Romdhane et al. | An improved traffic signs recognition and tracking method for driver assistance system | |
| CN103914702A (en) | System and method for boosting object detection performance in videos | |
| CN105303191A (en) | Method and apparatus for counting pedestrians in foresight monitoring scene | |
| CN101859382A (en) | License plate detection and identification method based on maximum stable extremal region | |
| CN110723432A (en) | Garbage classification method and augmented reality equipment | |
| CN104134078B (en) | Automatic selection method for classifiers in people flow counting system | |
| Momin et al. | Vehicle detection and attribute based search of vehicles in video surveillance system | |
| CN104978567A (en) | Vehicle detection method based on scenario classification | |
| CN102254428A (en) | Traffic jam detection method based on video processing | |
| CN109829382A (en) | The abnormal object early warning tracing system and method for Behavior-based control feature intelligent analysis | |
| CN103971106A (en) | Multi-view human facial image gender identification method and device | |
| CN104463232A (en) | Density crowd counting method based on HOG characteristic and color histogram characteristic | |
| CN103106414A (en) | Detecting method of passer-bys in intelligent video surveillance | |
| CN105809713A (en) | Object tracing method based on online Fisher discrimination mechanism to enhance characteristic selection | |
| CN103679214A (en) | Vehicle detection method based on online area estimation and multi-feature decision fusion | |
| CN104123714A (en) | Optimal target detection scale generation method in people flow statistics | |
| Hou et al. | A cognitively motivated method for classification of occluded traffic signs | |
| Pavaskar et al. | Real-time vehicle-type categorization and character extraction from the license plates | |
| CN101826150A (en) | Head detection method and device and head detection and category judgment method and device | |
| Ilayarajaa et al. | Text recognition in moving vehicles using deep learning neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20130925 |