CN104036250B - Video pedestrian detection and tracking - Google Patents
Video pedestrian detection and tracking Download PDFInfo
- Publication number
- CN104036250B CN104036250B CN201410266099.XA CN201410266099A CN104036250B CN 104036250 B CN104036250 B CN 104036250B CN 201410266099 A CN201410266099 A CN 201410266099A CN 104036250 B CN104036250 B CN 104036250B
- Authority
- CN
- China
- Prior art keywords
- people
- mrow
- tracking
- target
- people target
- 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
Abstract
The present invention relates to a kind of video pedestrian detection and tracking.This method is that excellent local characteristic is respectively provided with time domain and spatial domain according to wavelet transformation, based on video frame-to-frame differences, utilize multi-scale wavelet characteristic, extract foreground moving object, rigid body consistency for the important component of human body and is had according to the number of people, by to different people head's target sample learning and training, foreground moving object in video scene is classified and detected, determine number of people target, and based on the otherness of number of people color character, using particle filter and dynamic tracking chain, the number of people is tracked.The inventive method does not need specific hardware supported and scene condition to constrain, and method is easy, flexibly, easily realizes.
Description
Technical field
The present invention relates to a kind of video pedestrian detection and tracking, for digital image analysis with understanding.Belong to intelligence
Technical field of information processing.
Background technology
With the rapid growth of urban population and increasingly complicated, Mass disturbance, riot, attack of terrorism etc. of town environment
Cities and towns happen suddenly social security events, drastically influence cities and towns public safety, and the generation often for the social security events that happen suddenly, very
In big degree with human body behavioral activity tight association.How human body behavioral activity is effectively determined, and to its abnormal, suspicious actions certainly
Dynamic identification, it will help Security Officer is timely, handles crisis rapidly, and safe precaution ability is substantially improved, harmonious, flat so as to build
The social environment of peace, oneself turns into an important topic of current international community.Effectively to determine human body behavioral activity, its key is
How effective detection is with tracking the pedestrian position in video scene.
Human body is as non-rigid, and metamorphosis is various, and easily blocks, and video scene change is complicated various, leads
Cause effective video pedestrian detection very difficult with tracking.Main method includes at present：(1) number of people curvature measuring and geometry are used
Signature tracking method, the method foundation number of people curvature measuring number of people, and according to the geometric properties of the number of people, to number of people target following, the party
Method easily confirms as the number of people for the object of similar curvature, and tracking effect is undesirable, and false drop rate is high；(2) based on particle filter
Detection and tracing, this method are only capable of tracking single target, and when target occlusion, it is big with losing possibility；(3) texture analysis
Method, the method computation complexity is high, and generalization ability is low, is calculated when especially tracking multiple destination objects time-consuming serious.
The content of the invention
It is an object of the invention to calculate complicated, reliability in time energy for existing video pedestrian detection and tracking
It is low, detection with tracking object it is single, and to dynamic scene change sensitivity, noise jamming greatly, difficulty meet human body behavioral activity and
When analysis with understand require, there is provided a kind of video pedestrian detection and tracking, according to the number of people be human body important component
And there is rigid body consistency, by different people head's target sample learning with training, to the foreground moving pair in video scene
As being classified and being detected, determine number of people target, and based on the otherness of number of people color character, using particle filter and dynamic with
Track chain, realize effective tracking to the number of people under numerous conditions.
To reach above-mentioned purpose, idea of the invention is that：Excellent office is respectively provided with time domain and spatial domain according to wavelet transformation
Portion's feature, based on video frame-to-frame differences, using multi-scale wavelet characteristic, foreground moving object is extracted, is human body according to the number of people
Important component and there is rigid body consistency, by different people head's target sample learning and training, in video scene
Foreground moving object classified and detected, number of people target is determined, and based on the otherness of number of people color character, using particle
Filtering and dynamic tracking chain, are tracked to the number of people.
Conceived according to foregoing invention, the present invention uses following technical proposals：
A kind of video pedestrian detection and tracking, it is characterised in that comprise the following steps that：
1) pedestrian detection and tracking image capturing system are started：Gather video image；
2) foreground moving object is split
By the current frame image and previous frame image subtraction of camera acquisition, prospect fortune is partitioned into using small wave converting method
Dynamic subject area；
3) sample learning and training；
4) number of people target detection；
5) number of people target following；
6）Pedestrian's identity coherence confirms.
Above-mentioned steps 2) concrete operation step it is as follows：
（1）Current frame imageI t(x,y) and previous frame imageI t-1(x,y) subtract each other, obtain difference imageD(x,y)：
D(x,y)= I t(x,y)- I t-1(x,y)；
（2）Difference image multi-scale wavelet transformation：
；
Wherein,DFor difference image,h,vFilter operator in respectively horizontal, vertical direction,For convolution；
（3）The determination in foreground moving object region：Determine difference image multi-scale wavelet transformationEThreshold valueT 1, willEValue is high
InT 1All pixels composition region, be defined as foreground moving object region.
Above-mentioned steps 3) concrete operation step it is as follows：
（1）According to step 2）, the number of people Haar features of collection different human body Moving Objects, the number of composition number of people training sample
According to setD i =H i , and Haar features of human body limb and trunk, form the tag set of the non-number of peopleC i =T i }；
（2）Selection sort device, to above-mentioned data acquisition systemD i And tag setC i The sample set of composition（D i ,C i ）Supervised
Educational inspector practises, and adjusts parameter in grader, classifying quality is reached optimal.
Above-mentioned steps 4) concrete operation step it is as follows：
（1）According to step 2), the Haar features of foreground moving object are gathered, form test data setAD i =AH i }；
（2）Grader and its parameter according to determined by step 3), to test data setAD i Discriminant classification is carried out, really
Determine number of people target.
Above-mentioned steps 5) concrete operation step it is as follows：
(1) color space conversion：By the red of RGB color spaceR, it is greenG, it is blueBThree-component, determine the color of HSV color spaces
Adjust componentH, saturation degree componentSAnd luminance componentV：
Wherein,
V = max(R, G, B)
(2) number of people feature histogram is built：The number of people target according to determined by step 4), using in HSV color spaces
Chrominance componentH, saturation degree componentS, establish each componentmThe color histogram of level, and using the brightness in HSV color spaces
ComponentV, establishnLevel gray gradient histogram, and then establish confluent colours (H、S) histogramWith brightness (V) gray scale ladder
Spend histogramNumber of people feature histogramq r ：
Wherein,CFor normalization coefficient,
(3) number of people target following：The number of people feature histogram built according to step (2)q r , using particle filter, to field
Number of people target following in scape.
Above-mentioned steps 6) concrete operation step it is as follows：
(1) dynamic tracking chain is set：It is if tracking number of people target in scenen, the number of people target that is tracked according to step 5),
Dynamic tracking chain is setT i (i=1,…,n)；
(2) the distance between dynamic tracking chain calculates：According to step (1), dynamic tracking chain is calculatedT i Between it is European away from
Fromd ij (i=1,…,n,j=1,...,n)；
(3) number of people target whether shadowing：According to step (2), if the distance between dynamic tracking chaind ij Less than basis
The certain threshold value for the number of people that step 5) is predicted present frameT 2, then hidden by the number of people target that dynamically tracking chain is tracked
Gear, conversely, number of people target occlusion terminates or in the absence of blocking；
(4) dynamic tracking chain and the incidence matrix of testing result are established：Chain is tracked according to the dynamic of step (1)T i (i=
1,…,n) result and the number of people object detection results according to step 4)H j (j=1,…,m,mFor the number of people number of detection), establish and close
Join matrixM ij = D(T i , H j ), whereinDFor distance metric operator；
(5) incidence matrix minimum value calculates：According to incidence matrix determined by step (4)M ij , determine the minimum of the matrix
ValueD m (I is not equal to j)；
(6) relational matrix is built：According to minimum value determined by step (5)D m , obtain and whether associated with dynamic tracking chain
Relational matrixR j ：
(7) number of people detection is merged with tracking result：According to relational matrix determined by step (6)R j IfR j <1, then currently
Nobody head's mark is associated with dynamic tracking chain, shows that unmanned head's mark or the number of people target previously tracked have been left in scene
Scene；IfR j =1, then the number of people target in present frame and the dynamic tracking chain associate, use weight forw 1(0< w 1<1) detection
As a result it is with weightw 2(w 2=1- w 1) tracking result, obtain current number of people target location；IfR j >1, show to deposit in present frame
Associated in multiple number of people targets with dynamic tracking chain, now using number of people feature histogramq r Number of people target differentiation is carried out, and
Weight, which is respectively adopted, isw 1(0< w 1<1) testing result is with weightw 2(w 2=1- w 1) tracking result, it is determined that current each
Number of people target location.
The principle of the present invention is as follows：In the inventive solutions, it is respectively provided with according to wavelet transformation in time domain and spatial domain
Excellent local characteristic, based on video frame-to-frame differences, using multi-scale wavelet characteristic, foreground moving object is extracted, according to the number of people
Important component for human body and there is rigid body consistency, by different people head's target sample learning and training, to regarding
Foreground moving object in frequency scene is classified and detected, and determines number of people target, and based on the otherness of number of people color character,
Using particle filter and dynamic tracking chain, the number of people is tracked.
If at a certain moment, adjacent two field pictures are obtained respectivelyf(t n-1, x, y),f(t n , x, y), by two images by
Pixel seeks difference, obtains difference imageDiff(x, y)：
Wherein,DiffR,DiffG,DiffBDifference image red, green, blue three-component is corresponded to respectively, |f| it isfAbsolute value.
Based on above-mentioned neighbor frame difference, using wavelet transformation, foreground moving object region is partitioned into.According to two dimensional imageI（x,y）In yardstick 2 j WithkWavelet transformation on direction：
Then existx,yWavelet function on direction is represented by：
In formula,For smoothing filter function.
Thus it can determine that imageI(x, y) through functionAfter smothing filtering, the wavelet transformation under different scale is：
If gradient amplitudeReach local maximum along following gradient direction, then the point in image (x, y) it is multiple dimensioned side
Edge point
Accordingly, it may be determined that the marginal point under different scale.Because noise is sensitive to dimensional variation, therefore, sought using above-mentioned
Seek local amplitude maximum, it is impossible to effectively suppress noise.Effectively to overcome this influence, it is higher than certain threshold by seeking gradient amplitude
Value method, replacement seek local amplitude maximum, determine the marginal point of different scale.
Wherein,H, vFilter operator in respectively horizontal, vertical direction,TFor threshold value,For convolution operator.
Consider number of people feature mode spaceX, comprisingmIndividual patternx i Training set and corresponding class label, and it is false
It is set to two class classification problems.In each layerkIn, the importance of sample uses weight setD k (i) reflection, and meet。
In two classification problems, the study of Weak Classifier makes object functionMinimize：
Wherein, P [] is the empirical probability based on training sample observation.
Weight is updated as the following formulaD k (i)：
Wherein,Z k For normalization factor, meet,
Final grader is by owningkIndividual Weak Classifier is considering its weightRear weight a majority of the votes cast determines.
The number of people target location based on determined by above-mentioned grader, using the chrominance component in HSV color spacesH, saturation
Spend componentS, establish each componentmThe color histogram of level, and using the luminance component in HSV color spacesV, establishnLevel
Gray gradient histogram.On this basis, establish confluent colours (H、S) histogramWith brightness (V) shade of gray Nogata
FigureNumber of people feature histogramq r ：
Wherein,CFor normalization coefficient,
IfX t ,Z t RespectivelytThe number of people dbjective state and observation at moment, then after number of people tracking problem being converted into solution
Test probabilityp(X t |Z 1:t ), wherein,Z 1:t =(Z 1,…,Z t ) be totThe owner head obtained untill moment marks observation.
Using one group of particle for carrying weightClose approximation posterior probabilityp(X t |Z 1:t ), wherein,
For particle, possible number of people dbjective state is represented,For the weight of particle.
New particle is produced by resampling function, and the functional dependence is in number of people dbjective state and observation, i.e.,
New particle is updated using following weight：
And new particle is produced by following state transition function：
X t =F t (X t-1,U t )
Wherein,U t For system noise,F t For people head's target motion state.
To make above-mentioned number of people testing result consistent with tracking result, caused occlusion issue during human motion is overcome
And keep blocking the identity coherence after terminating, it is if tracking number of people target in scenen, and dynamic tracking chain is setT i (i=
1,…,n), calculate dynamic tracking chainT i Between Euclidean distanced ij (i=1,…,n,j=1,...,n), ifd ij It is pre- less than present frame
Certain threshold value of the number of people is surveyed, the number of people target for showing now to be tracked by dynamic tracking chain is blocked, conversely, dynamic tracking chain
The number of people target occlusion tracked terminates or in the absence of blocking.
Chain is tracked based on dynamicT i (i=1,…,n) result and current all mark inspections number of people objective resultsH j (j=1,…,m,mFor the number of people number of detection), establish dynamic tracking chain and the incidence matrix of testing result dataM ij = D(T i , H j ), whereinDFor
Distance metric operator.
ByM ij , determine the minimum value of the matrixD m (I is not equal to j), and according toD m , obtain is as follows with dynamic tracking chain
The relational matrix of no association：
IfR j <1, then it is associated with dynamic tracking chain currently without number of people target, show unmanned head's mark or first in scene
The number of people target of preceding tracking has left scene；IfR j =1, then the number of people target in present frame and the dynamic tracking chain associate, use
Weight isw 1(0< w 1<1) testing result is with weightw 2(w 2=1- w 1) tracking result, obtain current number of people target location；
IfR j >1, show that multiple number of people targets in present frame be present associates with dynamic tracking chain, now using number of people feature histogramq r
Number of people target differentiation is carried out, and weight is respectively adopted and isw 1(0< w 1<1) testing result is with weightw 2(w 2=1- w 1) with
Track result, it is determined that current each number of people target location.
The present invention compared with prior art, has following obvious prominent substantive distinguishing features and remarkable advantage：This
Invention is respectively provided with excellent local characteristic in time domain and spatial domain according to wavelet transformation, more using small echo based on video frame-to-frame differences
Dimensional properties, foreground moving object is extracted, rigid body consistency for the important component of human body and is had according to the number of people, by right
Different people head's target sample learning and training, are classified and are detected to the foreground moving object in video scene, determine people
The head marks, and based on the otherness of number of people color character, using particle filter and dynamic tracking chain, the number of people is tracked, transported
Easy, flexible, easily realization is calculated, when solving video pedestrian detection with tracking, it is desirable to detect and tracking object is single, to dynamically
Scene changes are sensitive, noise jamming is big, computing is complicated, and specific hardware supported and scene condition constraint；Improve and regard
Frequency pedestrian detection and the robustness of tracking, the pedestrian detection being suitable under complex background condition and tracking.
Brief description of the drawings
Fig. 1 is the flowsheet of the inventive method.
Fig. 2 is the original current frame image of video of one embodiment of the invention.
Fig. 3 is the two-value foreground moving object area image being partitioned into Fig. 2 examples.
Fig. 4 is the foreground moving object area image being partitioned into Fig. 2 examples.
Fig. 5 is the number of people testing result in Fig. 2 examples（Rectangular box）.
Fig. 6 is the number of people tracking result in Fig. 2 examples.
Embodiment
Details are as follows for the preferred embodiments of the present invention combination accompanying drawing：
Embodiment one：
Referring to Fig. 1, this video pedestrian detection and tracking, it is characterised in that comprise the following steps that：
1) pedestrian detection and tracking image capturing system are started：Gather video image；
2) foreground moving object is split
By the current frame image and previous frame image subtraction of camera acquisition, using small wave converting method, prospect is partitioned into
Moving Objects region；
3) sample learning and training；
4) number of people target detection；
5) number of people target following；
6）Pedestrian's identity coherence confirms.
Embodiment two：
The original current frame image of this example to the image shown in Fig. 2 as shown in Fig. 2 carry out consecutive frame difference, and carry out small
Ripple multi-scale transform, foreground moving object segmentation is carried out, obtain two-value foreground moving object region as shown in figure 3, according to the number of people
Important component for human body and there is rigid body consistency, by different people head's target sample learning and training, to regarding
Foreground moving object in frequency scene is classified and detected, and determines number of people target, and based on the otherness of number of people color character,
Using particle filter and dynamic tracking chain, the number of people is tracked；Concrete operation step is as follows：
1) pedestrian detection and tracking image capturing system are started：Gather video image；
2) foreground moving object is split：Concrete operation step is as follows：
（1）By camera acquisition such as Fig. 2 current frame imageI 1(x, y) and previous frame imageI 2(x, y) subtract each other, obtain
To difference imageD(x,y)：D(x,y)= I t(x,y)- I t-1(x,y)；
（2）Difference image multi-scale wavelet transformation：
；
Wherein,DFor difference image,h,vFilter operator in respectively horizontal, vertical direction,For convolution；
（3）The determination in foreground moving object region：Determine difference image multi-scale wavelet transformationEThreshold valueT, willEValue is high
InTAll pixels composition region, be defined as foreground moving object region.
Fig. 3 is the two-value foreground moving object region through above-mentioned gained, and Fig. 4 is the foreground moving object being partitioned into.
3) sample learning and training：The number of people Haar features of different human body Moving Objects are gathered, form number of people training sample
Data acquisition systemD i =H i , and Haar features of human body limb and trunk, form the tag set of the non-number of peopleC i =T i , adopt
With SVMs and Radial basis kernel function is selected, to above-mentioned data acquisition systemD i And tag setC i The sample set of composition（D i ,C i ）Learnt and trained, constantly change the penalty factor parameter in Radial basis kernel functionγ, correct recognition rata is reached most
It is high；
4) number of people target detection：To the foreground moving object shown in Fig. 3, Haar features are gathered, form test data setAD i =AH i , using fixed penalty factor parameterγThe support vector cassification based on Radial basis kernel function is carried out to differentiate,
Determine number of people target.Rectangular box in Fig. 5 show the number of people position through above-mentioned gained；
5) number of people target following
Concrete operation step is as follows：
(1) color space conversion：By the red of RGB color spaceR, it is greenG, it is blueBThree-component, determine the color of HSV color spaces
Adjust componentH, saturation degree componentSAnd luminance componentV：
Wherein,
V = max(R, G, B)
(2) number of people feature histogram is built：To the number of people target of Fig. 5 examples, using the chrominance component in HSV color spacesH, saturation degree componentS, establishHWithSThe color histogram that each 8 grades of component, and using the luminance component in HSV color spacesV, establish 8 grades of gray gradient histograms, and then establish confluent colours (H、S) histogramWith brightness (V) shade of gray is straight
Fang TuNumber of people feature histogramq r ：
Wherein,CFor normalization coefficient,
(3) number of people target following：According to constructed number of people feature histogramq r , using particle filter, in scene
Number of people target following.Number of people upper values in Fig. 6 are through the number of people sequence number obtained by above-mentioned tracking.
6）Pedestrian's identity coherence confirms
Concrete operation step is as follows：
(1) dynamic tracking chain is set：If tracking number of people target 2 in scene, the number of people target tracked according to Fig. 5 examples,
Dynamic tracking chain is setT i (i=1,2)；
(2) the distance between dynamic tracking chain calculates：According to step (1), dynamic tracking chain is calculatedT i Between it is European away from
Fromd ij (i=1,2,j=1,2)；
(3) number of people target whether shadowing：According to step (2), dynamic tracks the distance between chaind ij Not less than basis
The 75% of the number of people size that step 5) is predicted present frame, show that number of people target occlusion terminates or in the absence of blocking；
(4) dynamic tracking chain and the incidence matrix of testing result are established：Chain is tracked according to the dynamic of step (1)T i (i=1,
2) number of people testing result of result and Fig. 4 examplesH j (j=1,2), incidence matrix is establishedM ij = D(T i , H j ), whereinDFor Euclidean away from
Accorded with from metric operations；
(5) incidence matrix minimum value calculates：According to incidence matrix determined by step (4)M ij , determine the minimum of the matrix
ValueD m (I is not equal to j)；
(6) relational matrix is built：According to minimum value determined by step (5)D m , obtain and whether associated with dynamic tracking chain
Relational matrixR j ：
(7) number of people detection is merged with tracking result：According to relational matrix determined by step (6)R j =2, it is special using the number of people
Levy histogramq r Number of people target differentiation is carried out, and the testing result that weight is 0.5 and the tracking knot that weight is 0.5 is respectively adopted
Fruit, it is determined that current each number of people target location.
Claims (5)
1. a kind of video pedestrian detection and tracking, it is characterised in that comprise the following steps that：
1) pedestrian detection and tracking image capturing system are started：Gather video image；
2) foreground moving object is split
By the current frame image and previous frame image subtraction of camera acquisition, using small wave converting method, foreground moving is partitioned into
Subject area；
3) sample learning and training；
4) number of people target detection；
5) number of people target following；
6) pedestrian's identity coherence confirms, concretely comprises the following steps：
(1) dynamic tracking chain is set：If it is n that number of people target is tracked in scene, the number of people target tracked according to step 5), set
Dynamic tracking chain Ti, i=1 ..., n；
(2) the distance between dynamic tracking chain calculates：According to step (1), dynamic tracking chain T is calculatediBetween Euclidean distance dij,
I=1 ..., n, j=1 ..., n；
(3) number of people target whether shadowing：According to step (2), if the distance between dynamic tracking chain dijLess than according to step
5) the certain threshold k for the number of people predicted present frame, the then number of people target tracked by dynamic tracking chain are blocked, instead
It, number of people target occlusion terminates or in the absence of blocking；
(4) dynamic tracking chain and the incidence matrix of testing result are established：Chain T is tracked according to the dynamic of step (1)i, i=1 ..., n
And the number of people object detection results H according to step 4) as a resultj, j=1 ..., m, m is the number of people number of detection, establishes incidence matrix Mij
=D (Ti,Hj), wherein D is distance metric operator；
(5) incidence matrix minimum value calculates：According to incidence matrix M determined by step (4)ij, determine the minimum value D of the matrixm,
i≠j；
(6) relational matrix is built：According to minimum value D determined by step (5)m, obtain the relation whether associated with dynamic tracking chain
Matrix Rj：
(7) number of people detection is merged with tracking result：According to relational matrix R determined by step (6)jIf Rj<1, then currently without
Number of people target is associated with dynamic tracking chain, shows that unmanned head's mark or the number of people target previously tracked have left field in scene
Scape；If Rj=1, then the number of people target in present frame and the dynamic tracking chain associate, use weight as w1, 0<w1<1 detection knot
Fruit and weight are w2, w2=1-w1Tracking result, obtain current number of people target location；If Rj>1, show in present frame in the presence of more
Personal head's mark associates with dynamic tracking chain, now using number of people feature histogram qrNumber of people target differentiation is carried out, and respectively
Weight is used as w1, 0<w1<1 testing result and weight are w2, w2=1-w1Tracking result, it is determined that current each number of people target
Position.
2. video pedestrian detection according to claim 1 and tracking, it is characterised in that the step 2) foreground moving
The concrete operation step of Object Segmentation is as follows：
(1) current frame image It(x, y) and previous frame image It-1(x, y) subtracts each other, and obtains difference image D (x, y)：
D (x, y)=It(x,y)-It-1(x,y)；
(2) difference image multi-scale wavelet transformation：
<mrow>
<mi>E</mi>
<mo>=</mo>
<msqrt>
<mrow>
<msup>
<mrow>
<mo>(</mo>
<mi>D</mi>
<mo>&CircleTimes;</mo>
<mi>h</mi>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
<mo>+</mo>
<msup>
<mrow>
<mo>(</mo>
<mi>D</mi>
<mo>&CircleTimes;</mo>
<mi>v</mi>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
</mrow>
</msqrt>
<mo>;</mo>
</mrow>
Wherein, D is difference image, and h, v are respectively horizontal, the filter operator in vertical direction,For convolution；
(3) determination in foreground moving object region：Determine difference image multi-scale wavelet transformation E threshold value T1, E values are higher than T1
All pixels composition region, be defined as foreground moving object region.
3. video pedestrian detection according to claim 1 and tracking, it is characterised in that the step 3) sample learning
It is as follows with the concrete operation step of training：
(1) according to step 2), the number of people Haar features of collection different human body Moving Objects, the data set of composition number of people training sample
Close Di={ Hi, and Haar features of human body limb and trunk, form the tag set C of the non-number of peoplei={ Ti}；
(2) selection sort device, to above-mentioned data acquisition system DiWith tag set CiSample set (the D of compositioni, Ci) exercise supervision
Practise, and adjust parameter in grader, classifying quality is reached optimal.
4. video pedestrian detection according to claim 1 and tracking, it is characterised in that the step 4) number of people target
The concrete operation step of detection is as follows：
(1) according to step 2), the Haar features of foreground moving object is gathered, form test data set ADi={ AHi}；
(2) grader and its parameter according to determined by step 3), to test data set ADiDiscriminant classification is carried out, determines the number of people
Target.
5. video pedestrian detection according to claim 1 and tracking, it is characterised in that the step 5) number of people target
The concrete operation step of tracking is as follows：
(1) color space conversion：By the red R of RGB color space, green G, blue B three-components, determine that the tone of HSV color spaces divides
Measure H, saturation degree component S and luminance component V：
Wherein,
<mrow>
<mi>S</mi>
<mo>=</mo>
<mn>1</mn>
<mo>-</mo>
<mn>3</mn>
<mfrac>
<mrow>
<mi>min</mi>
<mrow>
<mo>(</mo>
<mi>R</mi>
<mo>,</mo>
<mi>G</mi>
<mo>,</mo>
<mi>B</mi>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mo>(</mo>
<mi>R</mi>
<mo>+</mo>
<mi>G</mi>
<mo>+</mo>
<mi>B</mi>
<mo>)</mo>
</mrow>
</mfrac>
</mrow>
V=max (R, G, B)
(2) number of people feature histogram is built：The number of people target according to determined by step 4), using the tone in HSV color spaces
Component H, saturation degree component S, establish the color histogram of each component m levelsAnd using the luminance component in HSV color spaces
V, establish n level gray gradient histogramsAnd then establish confluent colours (H, S) histogramIt is straight with brightness (V) shade of gray
Fang TuNumber of people feature histogram qr：
<mrow>
<msup>
<mi>q</mi>
<mi>r</mi>
</msup>
<mo>=</mo>
<msubsup>
<mi>Cq</mi>
<mi>c</mi>
<mrow>
<mi>m</mi>
<mo>&times;</mo>
<mi>m</mi>
</mrow>
</msubsup>
<msubsup>
<mi>q</mi>
<mi>v</mi>
<mi>n</mi>
</msubsup>
<mo>,</mo>
<mi>r</mi>
<mo>=</mo>
<mn>0</mn>
<mo>,</mo>
<mo>...</mo>
<mo>,</mo>
<msup>
<mi>m</mi>
<mn>2</mn>
</msup>
<mo>&times;</mo>
<mi>n</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
Wherein, C is normalization coefficient,
(3) number of people target following：The number of people feature histogram q built according to step (2)r, using particle filter, in scene
Number of people target following.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410266099.XA CN104036250B (en) | 2014-06-16 | 2014-06-16 | Video pedestrian detection and tracking |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410266099.XA CN104036250B (en) | 2014-06-16 | 2014-06-16 | Video pedestrian detection and tracking |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104036250A CN104036250A (en) | 2014-09-10 |
| CN104036250B true CN104036250B (en) | 2017-11-10 |
Family
ID=51467016
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410266099.XA Expired - Fee Related CN104036250B (en) | 2014-06-16 | 2014-06-16 | Video pedestrian detection and tracking |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104036250B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| RU2796096C1 (en) * | 2022-05-13 | 2023-05-17 | Акционерное общество "Научно-Производственный Комплекс "Альфа-М" | Method of tracking objects |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105005773A (en) * | 2015-07-24 | 2015-10-28 | 成都市高博汇科信息科技有限公司 | Pedestrian detection method with integration of time domain information and spatial domain information |
| CN105260712B (en) * | 2015-10-03 | 2019-02-01 | 上海大学 | A kind of vehicle front pedestrian detection method and system |
| CN106447694A (en) * | 2016-07-28 | 2017-02-22 | 上海体育科学研究所 | Video badminton motion detection and tracking method |
| CN106875425A (en) * | 2017-01-22 | 2017-06-20 | 北京飞搜科技有限公司 | A kind of multi-target tracking system and implementation method based on deep learning |
| CN107220629B (en) * | 2017-06-07 | 2018-07-24 | 上海储翔信息科技有限公司 | A kind of method of the high discrimination Human detection of intelligent automobile |
| JP6973258B2 (en) * | 2018-04-13 | 2021-11-24 | オムロン株式会社 | Image analyzers, methods and programs |
| CN111291599A (en) * | 2018-12-07 | 2020-06-16 | 杭州海康威视数字技术股份有限公司 | Image processing method and device |
| CN110889351B (en) * | 2019-11-18 | 2023-09-26 | 中国科学院深圳先进技术研究院 | Video detection method, device, terminal equipment and readable storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101216940A (en) * | 2008-01-08 | 2008-07-09 | 上海大学 | Video foreground moving object subdivision method based on wavelet multi-scale transform |
| CN102799935A (en) * | 2012-06-21 | 2012-11-28 | 武汉烽火众智数字技术有限责任公司 | Human flow counting method based on video analysis technology |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6785402B2 (en) * | 2001-02-15 | 2004-08-31 | Hewlett-Packard Development Company, L.P. | Head tracking and color video acquisition via near infrared luminance keying |
-
2014
- 2014-06-16 CN CN201410266099.XA patent/CN104036250B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101216940A (en) * | 2008-01-08 | 2008-07-09 | 上海大学 | Video foreground moving object subdivision method based on wavelet multi-scale transform |
| CN102799935A (en) * | 2012-06-21 | 2012-11-28 | 武汉烽火众智数字技术有限责任公司 | Human flow counting method based on video analysis technology |
Non-Patent Citations (2)
| Title |
|---|
| Spatio-temporal motion-based foreground segmentation and shadow suppression;Y.P.Guan;《The Institution of Engineering and Technology》;20101231;第4卷(第1期);正文第1-2节 * |
| 基于头肩轮廓特征的人头检测系统的研究;顾炯;《中国优秀硕士学位论文全文数据库 信息科技辑》;20120615;第2012年卷(第6期);正文第1.1节、2.1.1节、2.2节、4.4节、图4.14 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| RU2796096C1 (en) * | 2022-05-13 | 2023-05-17 | Акционерное общество "Научно-Производственный Комплекс "Альфа-М" | Method of tracking objects |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104036250A (en) | 2014-09-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104036250B (en) | Video pedestrian detection and tracking | |
| CN109583342B (en) | Human face living body detection method based on transfer learning | |
| CN107016357B (en) | Video pedestrian detection method based on time domain convolutional neural network | |
| Bertini et al. | Multi-scale and real-time non-parametric approach for anomaly detection and localization | |
| CN104899866B (en) | A kind of intelligentized infrared small target detection method | |
| CN103824070B (en) | A kind of rapid pedestrian detection method based on computer vision | |
| CN102982313B (en) | The method of Smoke Detection | |
| CN112016445B (en) | Monitoring video-based remnant detection method | |
| Zhang et al. | Region of interest extraction in remote sensing images by saliency analysis with the normal directional lifting wavelet transform | |
| CN108446690B (en) | Human face in-vivo detection method based on multi-view dynamic features | |
| CN113536972B (en) | Self-supervision cross-domain crowd counting method based on target domain pseudo label | |
| Asokan et al. | Machine learning based image processing techniques for satellite image analysis-a survey | |
| Mahlisch et al. | A multiple detector approach to low-resolution FIR pedestrian recognition | |
| Zhang et al. | Object detection/tracking toward underwater photographs by remotely operated vehicles (ROVs) | |
| CN106447694A (en) | Video badminton motion detection and tracking method | |
| CN105303571A (en) | Time-space saliency detection method for video processing | |
| Zhu et al. | Fast detection of moving object based on improved frame-difference method | |
| CN104899559B (en) | A kind of rapid pedestrian detection method based on video monitoring | |
| Anderson et al. | Algorithm fusion in forward-looking long-wave infrared imagery for buried explosive hazard detection | |
| Xu et al. | COCO-Net: A dual-supervised network with unified ROI-loss for low-resolution ship detection from optical satellite image sequences | |
| Chang et al. | Locating waterfowl farms from satellite images with parallel residual u-net architecture | |
| Chen et al. | A novel AMS-DAT algorithm for moving vehicle detection in a satellite video | |
| Liu et al. | An automatic high confidence sets selection strategy for SAR images change detection | |
| Hu et al. | Anomaly detection in crowded scenes via sa-mhof and sparse combination | |
| Chen et al. | Detection and adaptive video processing of hyperopia scene in sports video |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| 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: 20171110 Termination date: 20200616 |