CN106408594B - Video multi-target tracking based on more Bernoulli Jacob's Eigen Covariances - Google Patents
Video multi-target tracking based on more Bernoulli Jacob's Eigen Covariances Download PDFInfo
- Publication number
- CN106408594B CN106408594B CN201610860087.9A CN201610860087A CN106408594B CN 106408594 B CN106408594 B CN 106408594B CN 201610860087 A CN201610860087 A CN 201610860087A CN 106408594 B CN106408594 B CN 106408594B
- Authority
- CN
- China
- Prior art keywords
- target
- indicate
- moment
- particle
- targets
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of video multi-target trackings based on more Bernoulli Jacob's Eigen Covariances, it belongs to artificial intelligence and intelligent information processing technology field, and it is inaccurate close to, change in size and tracking mainly to solve the problems, such as during number under complex environment is unknown and the video multi-target tracking of variation that there are targets.This method combines multiple features covariance technique by under more Bernoulli Jacob's filter frames, introducing integrogram thought, is realized using particle filter method and is tracked to the video multi-target of number of variations；On this basis, propose that target close to adaptation mechanism and target size adaptation mechanism, is realized respectively to the self-adaptive processing close to target and track window；It finally uses particle label method to realize to track the movement locus self-adapting estimation of video multi-target.The present invention has stronger robustness and anti-interference ability, can meet the design requirement of Practical Project system, has good engineering application value.
Description
Technical field
The invention belongs to intelligent information processing technology fields, are related to the visual frequency multi-object tracking method of parameter.Specifically
It is the video multi-target tracking of a kind of feature based covariance matrix and more Bernoulli Jacob filtering, can be used for various traffic pipes
Video multi-target detection in the systems such as system, robot navigation and video monitoring and tracking.
Background technology
In computer vision application field, target numbers variation and intersect, close to etc. video multi-targets tracking be one non-
Often important and challenging problem is always the hot and difficult issue in video tracking research.Especially when target intersect or
Close to when, it is easy to there is target tracking accuracy decline, or even leakage is with directly affecting subsequent target identification, classification, behavior point
The processing such as analysis.
Currently, being directed to video monotrack, main method has average drifting, Kalman filtering, particle filter and spy
Sign matching etc..Wherein, how to choose suitable target signature to describe target comprehensively, be a key in target following
The quality of step, goal description will also directly affect subsequent tracking accuracy.In recent years, the scholars such as Tuzel are in paper《Region
covariance:a fast descriptor for detection and classification》Middle proposition covariance is calculated
Son not only reduces complexity, but also have by multiple Fusion Features of target a to covariance matrix than union feature
Stronger goal description ability, advantageously accounts for the interference problem in video frequency object tracking.
For multiple target tracking, early stage mainly uses data association technique, and calculation amount is larger, and can only track number
Know and fixed multiple target.In recent years, stochastic finite collection (Random Finite Set, RFS) is theoretical has obtained the extensive of scholar
Concern, in the case that do not need complex data it is associated can effectively track number it is unknown and variation multiple target.Especially
The more Bernoulli Jacob of gesture equilibrium multiple target that the scholars such as Mahler propose filter (CBMeMBer) method, by calculating missing inspection target and amount
Newer more Bernoulli Jacob's stochastic finite collection are surveyed, the posterior probability density function of directly approximate entire multiple target state set both can be with
The tracking accuracy for ensureing algorithm, also reduces calculating cost, unknown in number and variation point target and video frequency object tracking neck
Domain is widely applied.But when target is close to being blocked, these methods are difficult to differentiate between multiple targets, and target is caused to be leaked
Estimation；In addition, when significant change occur in target sizes, track window tends not to adaptively, tracking accuracy be caused to be decreased obviously.
Invention content
To the above problem, the present invention is under more Bernoulli Jacob's filter frames, introduced feature covariance technique and particle filter skill
Art proposes a kind of visual frequency multi-object tracking method of adaptive parameter, to be difficult to standard when solving target close to change in size
Really the problem of tracking, and it is further introduced into particle label technology on the basis of this method, video object is respectively transported with realizing
Dynamic rail mark accurately tracks, and improves adaptivity and robustness that the method for the present invention tracks video multi-target.
Realizing the key technology of the present invention is：During being tracked to video multi-target, frame is filtered in more Bernoulli Jacob first
Introduced feature covariance operator describes the appearance of each video object under frame, effectively to inhibit background interference；Then, tight to target
Adjacent part is analyzed, and is proposed target close to self-adaptive processing mechanism according to target area situation of change and is tracked at window adaption
Reason mechanism improves target numbers tracking accuracy and improves the adaptive ability of track window to efficiently separate close to target；Finally,
Using particle filter technology and particle label technology, video multi-target is continuously tracked in realization.
To realize above-mentioned target, steps are as follows for specific implementation：
(1) initialization step：
(1a) initial time k=0, video totalframes are N, and initial target state set isInitialized target shape
StateWhereinIndicate initial target i rectangle frame top left co-ordinates,Indicate initial target i squares
The width of shape frame,Indicate the height of initial target i rectangle frames, M0Indicate initial target number, initial target existing probability Ps=
0.99, calculate the Eigen Covariance of initial target iAnd carry out particle initialization sampling；
(1b) initializes newborn dbjective state collectionWherein It indicates
The coordinate in the newborn target i rectangle frames upper left corner,Indicate the width of newborn target i rectangle frames,Indicate newborn target i rectangle frames
Height, MΓIndicate newborn target number；The hypothesis on location of newborn target is fixed in certain coordinate range in the present invention, and is assumed newborn
The existing probability of target is 0.02；The Eigen Covariance of newborn target i is expressed asAnd according to newborn parameter initialization new life
The sampling particle of target；
(1c) initialization sampling Fe coatings, particle maximum number are Lmax, particle minimal amount is Lmin, initial target shape
State covariance Σ0=diag (1,1,0.1,0.1)；
(2) dbjective state prediction steps：
(2a) assumes that target movement model is random walk model, i.e.,：
X (k+1)=x (k)+e (k)
Wherein, x (k) indicates that the state of k moment targets, e (k) indicate k moment zero mean Gaussian white noises, dbjective state association
Variance is
(2b) target prediction, it is assumed that more Bernoulli parameter set representations can be used in the posterior probability density at k-1 moment, multiple target
For：
Wherein,WithThe existing probability and probability distribution of k-1 moment targets i, M are indicated respectivelyk-1Indicate the k-1 moment
Target number；Multiple target probability density after then predicting is represented by：
Wherein,Indicate that the k moment survives more Bernoulli parameter collection of target,WithPoint
The existing probability of target i that Biao Shi not survive at the k moment and prediction probability are distributed,Indicate k moment new life targets
More Bernoulli parameter collection,WithThe existing probability and probability distribution of k moment new life targets i, M are indicated respectivelyΓ,kWhen indicating k
Carve newborn target number；
(3) target likelihood calculates step：
(3a) present invention can handle grayscale image sequence and color image sequence, wherein five Wei Te of gray level image extraction
Sign is respectively：Gray scale, the First-order Gradient in the directions m and n, second order gradient；Coloured image extraction three-dimensional feature be respectively：Cromogram
Tri- color value of H, S, V of picture；
The Eigen Covariance T of initialized target template and the Eigen Covariance matrix F of candidate target, and calculate two covariances
The similarity measure of matrix, i.e.,：
D (T, F)=| | log (T)-log (F) | |
(3b) avoids target by partial occlusion, enhances the resolution energy to target in order to which more accurately target is described
Object representation is 5 piecemeals by power, the present invention, calculates the Eigen Covariance matrix of each block；And a fusion process is used, melt
The similarity measure for closing all pieces ignores the smallest blocks of similarity measure in candidate target piecemeal, obtains new overall similarity
Estimate D, i.e.,：
Wherein, min expressions are minimized function,Indicate the feature association side of the ξ piecemeal of k moment target i templates
Difference,Indicate the Eigen Covariance of the ξ piecemeal of k moment candidate targets i；
After (3c) acquisition overall similarity estimates D, the likelihood of the sampling particle j of k moment candidate targets i is expressed as：
Wherein, [10,30] parameter lambda ∈；
In addition, the feature templates in the present invention are calculated according to the composite character covariance matrix at preceding Φ moment,
I.e.：
Wherein, Fτ, ξIt indicates in τ moment target state estimator statesThe ξ piecemeal Eigen Covariance；
(4) dbjective state updates step：
The k moment, it is assumed that the prediction probability density of video multi-target stochastic finite collection is represented by：
Then updated multiple target posterior probability density is represented by：
Wherein,Indicate the more Bernoulli parameter collection being updated to prediction in the case of missing inspection,Indicate the more Bernoulli parameter collection being updated to measuring collection, ZkIndicate the measurement at k moment
Set, z indicate to measure the element of set；Particle weights are calculated according to step (3)；
(5) dbjective state extraction step：
In order to avoid sample degeneracy problem, resampling is carried out to sampling particle collection, and according to the sampling updated power of particle
Value estimates current k moment dbjective state collection Xk, in addition, in order to avoid the redundancy of more Bernoulli parameter collection quantity, it will in the present invention
Target of the existing probability less than 0.01 is deleted；
(6) target is close to adaptive step：
When two mesh close to when, two target discriminations are often a target by tracking result, to cause tracking wrong
Accidentally, or even there is target leakage with phenomenon；For such case, the present invention is proposed close to self-adaptive processing mechanism；
(7) target following window adaption step：
When video object size varies widely, track window cannot include whole features of target, and information is caused to lose
It loses；For this purpose, the present invention proposes track window self-adaptive processing mechanism so that the variation of track window energy adaptive targets size；
(8) particle label step：
The present invention is by being marked particle, to identify each target, to realize the track following to target.
The present invention has the following advantages：
(1) more Bernoulli Jacob's filtering techniques are used in the present invention, it is more effectively realizes unknown to number and variation video
Target following；And due to using particle label technology, the movement locus of each target can be efficiently extracted；
(2) present invention becomes target jamming, illumination due to carrying out general description to target using multiple features covariance operator
Change and target deformation etc. have stronger robustness, and can effectively solve to cause under tracking accuracy since target is blocked
The problem of drop；
(3) present invention proposes that target following window mechanism and target close to adaptation mechanism, can effectively be distinguished close to target,
And realize the adaptive of target following window, to improve algorithm keeps track precision.
Description of the drawings
Fig. 1 is the overall flow figure of the present invention；
Fig. 2 is the integrogram thought schematic diagram in the method for the present invention；
Fig. 3 is the target segment schematic diagram of the calculating Eigen Covariance matrix in the method for the present invention；
Fig. 4 is target in the method for the present invention close to adaptive design sketch；
Fig. 5 is the target following window adaption design sketch in the method for the present invention；
Fig. 6 is the track window adaptive tracing comparative result figure of the method for the present invention and tradition PF-MeMBer methods；
Fig. 7 is the track window adaptive tracing result OSPA value comparison diagrams of the method for the present invention and tradition PF-MeMBer methods；
Fig. 8 is the cross jamming tracking result comparison diagram of the method for the present invention and tradition PF-MeMBer methods；
Fig. 9 is the cross jamming tracking result OSPA value comparison diagrams of the method for the present invention and tradition PF-MeMBer methods；
Figure 10 be the change number multiple target of the method for the present invention and tradition PF-MeMBer methods and block interference tracking result pair
Than figure；
Figure 11 is the change number multiple target of the method for the present invention and tradition PF-MeMBer methods and blocks interference tracking result
OSPA value comparison diagrams.
Specific implementation mode
One, basic theory introduction
Bernoulli Jacob's filtering principle more than 1.
More Bernoulli Jacob's stochastic finite collection X are represented byThat is M mutual indepedent single Bernoulli Jacob's stochastic finite collection X(i)Union, r(i)And p(i)Its existing probability and probability distribution respectively, then spatially the probability of more Bernoulli Jacob's stochastic finite collection is close
Degree is represented by：
Stochastic finite collection can be described by its probability density, and the average potential estimation of the set is target numbers estimation.
Assuming that a parameter setBernoulli Jacob's stochastic finite collection more than one can be described, then the more Bernoulli Jacob's filtering of multiple target
It is exactly that state set and observation collection are all used into more Bernoulli Jacob's stochastic finite collection approximate representations, passes through recursion r(i)And p(i)Realize more mesh
Mark tracking.Its algorithm steps is as follows：
(1) it predicts：
Assuming that at the k-1 moment, the posterior probability density of multiple target is represented by：
Multiple target probability density after then predicting is represented by：
Wherein,
<a,b>Indicate the dot product of variable a and b, fk|k-1(x |) and pS,kSingle goal state transition probability density is indicated respectively
Function and target survival probability,For more Bernoulli parameter collection of k moment new life targets.
(2) it updates：
The k moment, if the prediction probability density of multiple target stochastic finite collection is represented by：
Updated posteriority multiple target probability density is represented by：
Wherein,
ψk,z(x)=gk(z|x)pD,k(x)
Wherein, gk(z | x) it indicates to measure likelihood, pD,k(x) detection probability, Z are indicatedkIt indicates to measure set, κk(z) indicate miscellaneous
Wave density function.
2. Eigen Covariance matrix
If region R indicates that W × H ties up gray level image, FRIndicate the W × H × d dimensional feature figures extracted from R, i.e.,：
Wherein, φ () is the mapping function of feature extraction, and five dimensional feature of gray level image is respectively：Gray scale, the directions m and n
First-order Gradient, second order gradient.For arbitrary rectangular area R, the corresponding d × d dimensional features covariance matrix in the region can be used
It indicates：
Wherein, flFor the d dimensional feature vectors corresponding to first of pixel, fl∈FR, l=1,2 ..., θ, θ=W × H, μ
For the characteristic mean vector of region R：
The present invention is under more Bernoulli Jacob's filter frames, using particle filter realization video multi-target tracking, however in grain
In son filtering, candidate target is obtained by sampling particle, still, if individually calculating feature association side to each candidate target
Poor matrix, then the pixel of candidate target lap, which will be repeated, counts multiple, increases calculating cost.It is asked to solve this
Topic, present invention introduces integrogram (Integral image, II) thoughts to improve computational efficiency.
Integrogram is extended to by integral image, in integrogram, the integral image II of pixel (m', n') (m',
N' it) is defined as the sum of all pixels point in the rectangular area of this upper left side, i.e.,：
The integrogram of arbitrary size region A can be quickly calculated by the integral image on its four vertex, i.e. II (A)=II
(m ", n ")+II (m', n')-II (m ", n')-II (m', n "), as shown in Figure 2.
The average gray AVRG (A) of region A is：
Integrogram is suitable for counting a large amount of rectangular areas, and these rectangular areas have a large amount of overlapping region again, can drop
Low calculation amount.
Covariance matrix CRIn (i, j) element can be expressed as：
If note：
Pm,n=[P (m, n, 1) ..., P (m, n, d)]T
Wherein,
Then region covariance is represented by：
Wherein, θ=(m "-m') (n "-n').After building integrogram, the calculating of the covariance matrix in any region is complicated
Degree will be reduced to O (d2), and it is unrelated with area size.
Two, the present invention is based on video multi-target trackings under more Bernoulli Jacob's filter frames of Eigen Covariance
Referring to Fig.1, specific implementation step of the invention includes as follows：
Step 1. initialized target collection, newborn object set sample Fe coatings
(1.1) initial time k=0, video totalframes are N, and initial target state set isInitialized target
StateWhereinWithIndicate the coordinate in the initial target i rectangle frames upper left corner,Indicate initial target
The width of i rectangle frames,Indicate the height of initial target i rectangle frames, M0Indicate initial target number, initial target existing probability Ps=
0.99, calculate the Eigen Covariance of initial target iAnd carry out particle initialization sampling.
(1.2) initializing newborn dbjective state collection isWhereinWhereinWith
Indicate the coordinate in the newborn target i rectangle frames upper left corner,Indicate the width of newborn target i rectangle frames,Indicate newborn target i rectangles
The height of frame, MΓIndicate newborn target number.The newborn hypothesis on location of newborn target is fixed in certain coordinate range in the present invention, and
Each frame has target new life, existing probability 0.02.Calculate the Eigen Covariance of newborn target iThe newborn target of initialization
Sampling particle.
(1.3) initialization sampling Fe coatings, particle maximum number Lmax, particle minimal amount Lmin, initial target state
Covariance Σ0=diag (1,1,0.1,0.1).
Step 2. predicts dbjective state
(2.1) assume that target movement model is random walk model, i.e.,：
X (k+1)=x (k)+e (k)
Wherein, x (k) indicates that the state of k moment targets, e (k) are k moment zero mean Gaussian white noises, dbjective state association side
Difference is
(2.2) assume the k-1 moment, it is known that more Bernoulli Jacob's posterior densities areWherein,With
The existing probability and probability distribution of k-1 moment targets i are indicated respectively, and eachBy one group of weighted sample (particle)It indicates, i.e.,：
Wherein,Indicate the weights of j-th of sampling particle of k-1 moment targets i,Indicate k-1 moment targets i's
The state vector of j-th of sampling particle,Indicate that the sampling particle number of k-1 moment targets i, δ () are Dirac function.
(2.3) prediction of multiple target posterior probability density is expressed as：
Wherein,Indicate that the k moment survives more Bernoulli parameter collection of target,WithPoint
Wei not survive at the k moment existing probability of target i and the predicted value of probability distribution, Mk-1Indicate k-1 moment target numbers,Indicate more Bernoulli parameter collection of k moment new life targets,WithRespectively k moment new life target i's deposits
In probability and probability distribution, MΓ,kIndicate k moment new life target numbers；
During prediction, newborn particle is by newborn target componentDirectly sampling obtains, target of surviving
More Bernoulli parametersPredicted value be respectively：
Wherein, pS,kFor k moment target survival probabilities,
Step 3., which calculates, measures likelihood
(3.1) feature templates of current goal are selected.
The present invention can handle grayscale image sequence and color image sequence, wherein five dimensional features point of gray level image extraction
It is not：Gray scale, the First-order Gradient in the directions m and n, second order gradient；Coloured image extracts three-dimensional feature, i.e. coloured image HSV colors
Tri- color value of H, S, V in space.
For clarification of objective template of surviving：If the target is survived, duration is less than Φ moment, i.e. ks＜ Φ, then current mesh
Mark the feature templates in the dbjective state that feature templates are the previous moment Objective extraction；The duration k if the target is surviveds>=Φ,
The fusion feature template in dbjective state that then current goal feature templates extract for the Φ moment before the target, i.e.,：
Wherein, Fτ,ξIt indicates in τ moment estimated statesThe ξ piecemeal Eigen Covariance.
It is primary that newborn clarification of objective template only needs initialization to calculate, without recalculating every time.
(3.2) Eigen Covariance matrix and its similarity measurement are calculated.
The present invention uses the difference between logarithm-euclidian metric Eigen Covariance matrix, if target model features are assisted
Variance is T, and candidate target Eigen Covariance is F, then logarithm-Euclidean distance between T and F is represented by：
D (T, F)=| | log (T)-log (F) | |
In order to which more accurately target is described, avoids target by partial occlusion, enhance the resolution capability to target, originally
Invention models target using piecemeal Descriptive strategies.Object representation is 5 piecemeals by the method for the present invention, and each block is counted respectively
Its Eigen Covariance matrix is calculated, as shown in Figure 3.It calculates and corresponds to the similarity measure of each piecemeal between two images, and adopt
The similarity measure that all pieces are merged with a fusion process, ignores the smallest blocks of similarity measure in candidate target piecemeal, obtains
Estimate D to overall similarity, i.e.,：
Wherein, min expressions are minimized function,The similitude of the sampling particle j and target template that indicate target i are surveyed
Degree,Indicate the Eigen Covariance in the ξ piecemeal of k moment target i templates,Indicate the sampling of k moment candidate targets i
The Eigen Covariance of the ξ piecemeal of particle j.
(3.3) the measurement likelihood of the sampling particle j of target i is calculated.
In the similarity measure for sampling the particle j and target template for obtaining target iAfterwards, the sampling of k moment candidate targets i
The likelihood of particle j can be calculated by following formula：
Wherein, [10,30] parameter lambda ∈.
Step 4. updates dbjective state
The k moment, it is assumed that more Bernoulli Jacob's posterior densities after prediction areWherein,By one
Group weighting particleIt indicates, i.e.,：
Wherein,Indicate the predicted value of j-th of sampling particle weights of k moment targets i,Indicate k moment targets i
J-th of sampling particle state prediction, prediction Gaussian component number Mk|k-1=Mk-1+MΓ,k。
Then updated multiple target posterior probability density is represented by：
Wherein,Indicate the more Bernoulli parameter collection being updated to prediction in the case of missing inspection,Indicate the more Bernoulli parameter collection being updated to measuring collection, ZkIndicate the measurement collection at k moment
It closes, z indicates to measure the element of set；Mk=Mk|k-1+|Zk|, | Zk| indicate the number measured,WithThe k moment is indicated respectively
The update of the existing probability and probability distribution of target i, i.e.,：
Wherein, It indicates to measure ykParticle is predicted to the k momentLikelihood,
It can be obtained according to step (3.3).
Step 5. dbjective state is extracted
The problem of increasing in order to avoid sample degeneracy and Bernoulli Jacob's component carries out resampling to particle, and deletes in the presence of general
Rate is less than 0.01 Bernoulli Jacob's component, if existing probability is more than or equal to 0.5, extracts the state of current k moment target iI.e.
Step 6. is handled close to objective self-adapting
(6.1) judgement target whether close to
By horizontal direction for, it is assumed that target i and the horizontal distance of target j areIfLess than threshold valuesJudgement
It essentially coincides, need not handle on two target level directions.WhenMore than threshold valuesAnd it is less than threshold valuesWhen, judge two mesh
It is marked in horizontal direction close to and doing separating treatment, the location parameter of more fresh target.
(6.2) close to target separating treatment
wi、wjThe width for indicating the rectangle frame of target i, target j respectively, works as judgement
Close to rear on two target level directions, the step of carrying out horizontal separation processing to two targets, is as follows：
The state of (6.2.1) target i is xi=[mi,ni,wi,hi], the state of target j is xj=[mj,nj,wj,hj], α=
0.15；
The horizontal distance of (6.2.2) two targets isIfExecute step (6.2.3)；
Otherwise, without carrying out separating treatment；
(6.2.3) is if mi≤mj, then mi=mi-wiα,mj=mj+wjα；Otherwise, mi=mi+wiα,mj=mj-wjα；
Vertical direction close to judge, single-frame images target separating resulting identical as horizontal direction method as method for separating and processing
As shown in Figure 4.
(6.3) in the present invention, two larger targets of coincidence factor are merged, merging condition is：Liang Ge rectangular targets area
The lap area in domain is more than compared with the 80% of Small object region area.
The processing of step 7. target following window adaption
(7.1) target extended mode is obtained.
As shown in fig. 5, it is assumed that the estimated state of kth frame target i is expressed asWith rectangle frame L1Table
Show；The extended mode of target i estimated states is expressed asWith rectangle frame L2It indicates, λ is taken in the present invention
It is 1.4.
(7.2) target state estimator state updates.
By Fig. 5 (a) binaryzations, the rectangle frame L of target i is obtained2The interior minimum rectangle frame for including largest connected regionSuch as
Rectangle frame L in Fig. 5 (b)3, then target state estimator state be updated toIt is 0.6 that β is taken in the present invention.
(7.3) treatment on special problems.
In some cases, when two targets close to when, the connected region of two targets is merged into a connected region；Or
For person since background interference, a target are made of two or more connected regions, these situations may lead to largest connected region
It mutates.For the present invention by using the state fusion method at multiple moment, solving largest connected region mutagenesis leads to track window certainly
The problem of adjustment failure, i.e.,：
Wherein, μ=0.5, d=5 are taken.That is the state x of k moment targets ik,iBe for the preceding 5 frame target state with it is current when
It carvesWeighted sum.
Step 8. particle label technology
(8.1) label prediction.More Bernoulli Jacob's component particle labels at k-1 moment, target of surviving are represented by：
Wherein, Mk-1Indicate the number of Bernoulli Jacob's component,Indicate the population of i-th of Bernoulli Jacob's component, andThe particle label of i.e. same Bernoulli Jacob's component is identical.The particle of newborn more Bernoulli Jacob's components
Label is represented by：
Wherein, MΓIndicate the number of newborn Bernoulli Jacob's component,Indicate the population of i-th of newborn Bernoulli Jacob's component, andThen the prediction of Bernoulli Jacob's component label can be expressed as：Tk|k-1=Tk-1+TΓ。
(8.2) flag update.The label measured with new rear Bernoulli Jacob's component is represented by：
Wherein, | Zk| it indicates to measure number,
(8.3) resampling.Former generation's particle label having the same of particle and they after resampling, primary after resampling
Exert sharp component label be：
Wherein, MkTo measure with new rear Bernoulli Jacob's component number,
It can recognize that the movement locus of each target by comparing the label of each Bernoulli Jacob's component.Wherein, when target is closed
And when, the label for merging target is the label of two larger targets of target existing probability.
The effect of the present invention can be further illustrated by following experiment：
1. experiment condition and parameter
It is Intel Core i5-3470,3.2GHz that experiment, which is in processor, inside saves as the Dell computer platforms of 4GB
On, it is completed using 2010 simulation softwares of MATLAB.The data that experiment one and experiment two use all are from Terravic
The infrared video sequence image chosen in Research Infrared Database databases, resolution ratio is 320 × 240.
Experiment be mainly used for verifying the method for the present invention in the case that video multi-target close to, change in size and background and noise jamming with
Track performance.Experiment respectively will discussed the scholars such as inventive method and Hoseinnezhad R in terms of tracking accuracy and stability
Text《Visual tracking in background subtracted image sequences via multi-
Bernoulli filtering》The conventional particle of middle proposition filters more bernoulli methods (PF-MeMBer) and makees performance comparison point
Analysis.
The data that experiment three uses are the video datas of two-way driven automobile on the shooting road surface of laboratory, are color video sequence
Row image, resolution ratio are 901 × 401.Group experiment is mainly used for verifying the video that the method for the present invention is unknown to number and changes
Multiple target tracking performance, there are target is newborn and disappear and the case where partial target is blocked in the scene.
Also by multiple Monte-Carlo Simulation in experimentation, to inventive method in tracking accuracy, steady from statistical property
Qualitative and calculating time etc. is compared and analyzes.It is respectively mean error to select tracking performance evaluation criterionIt is average
Root-mean-square errorIt is average to lose with rateAnd average operating timeIt is respectively defined as：
Wherein,For the c times Monte Carlo experiment first of target of k moment pair estimated state,It is covered for the c times special
Caro experiment tracks the run time of first of target at the k moment, c=1 ..., C, k=1 ..., K, l=1 ..., L,For k when
The time of day of first of target is carved, C is Monte Carlo total degree in experiment, and K is image sequence totalframes in experiment, and L is experiment
Middle target total number, V are that C Monte Carlo experiment is lost with total degree, and mistake is more than really with being defined as the tracking error at the moment
25% pixel unit of dbjective state.
2. experiment and interpretation of result
The method of the present invention is compared with tradition PF-MeMBer methods, is mainly carried out in terms of following two：
Experiment 1：Target following window adaption
In this group of experiment scene, totally 150 frame, target size change greatly image sequence, and algorithm keeps track window is proposed with verification
Adaptive ability.Left side target is moved to the left, and target size reduces.Right side target moves right, and target size becomes larger.If
Track window cannot be adjusted adaptively, then for the target of size reduction, track window is interior will to increase background and clutter, target itself
Pixel proportion reduce, it is possible to cause target leakage with；For the target become large-sized, track window cannot include mesh in time
All characteristic informations are marked, information may be caused to lose.
Fig. 6 gives the tracking result of the method for the present invention and tradition PF-MeMBer methods, wherein Fig. 6 (a) is tradition PF-
MeMBer method tracking results, Fig. 6 (b) are the method for the present invention tracking result., it is apparent that the track window of the method for the present invention
It can adaptively be adjusted according to the variation of target size, and preferable tracking result can be provided, and traditional PF-MeMBer
The method that random value is used to the track window of particle in method, it can be seen that the target larger to change in size, track window is not
The variation for adapting to target causes target that cannot be fully extracted, as shown in the 112nd frame in Fig. 6 (a) and 150 frames.
Fig. 7 gives the comparison diagram of the OSPA distance statistics of this group experiment, it can be seen that the tracking essence of this hair inventive method
Degree will be apparently higher than traditional PF-MeMBer methods.
This group is tested and carries out 100 Monte Carlo simulations, the performance of tracking accuracy, stability and calculating time etc.
Evaluation is as shown in table 1.As can be seen that tradition PF-MeMBer methodsWithIt is higher, although without complete
Target is lost, but departure degree is larger；And the method for the present inventionWithIt is relatively low, it can realize to scale
Change greatly the adaptive and accurate tracking of target.Although there are it is a degree of lose with, be since right side target scale is larger,
Corresponding tracking error is also larger, has the error of several frames to be more than 25% pixel unit, but do not lose target.Run time
Also lower than traditional method, there is apparent tracking advantage.
Table 1 tracks the tracking performance evaluation in the case of window adaption
Experiment 2：Target cross jamming is tested
In this group of experiment scene, image sequence totally 100 frame, target carries out crisscross motion, experience close to, merge and separate
Etc. processes, by the group test to verify the method for the present invention to the process performance close to target, and to the flight path of cross-goal
Tracking performance.
Fig. 8 gives the tracking result of the method for the present invention and tradition PF-MeMBer methods, wherein Fig. 8 (a) is tradition PF-
MeMBer method tracking results, Fig. 8 (b) are the method for the present invention tracking result., it is apparent that the method for the present invention has preferably
Tracking performance, can be effectively treated close to target following, as shown in the 57th frame of Fig. 8 (b) and 68 frames, proposition method can have
Effect separates two close to target, improves tracking efficiency；In addition, by particle label method, the method for the present invention can also accurately be estimated
Count the track of crisscross motion target.
Fig. 9 gives the comparison diagram of the OSPA values of this group experiment, it can be seen that when target intersection interferes with each other, tradition
The error of PF-MeMBer methods is uprushed, and target leakage occurs with and the method for the present invention can be adaptive when cross jamming occurs in target
It handles with answering close to target, there is preferable tracking performance.
This group is tested and carries out 100 Monte Carlo simulations, the performance of tracking accuracy, stability and calculating time etc.
Evaluation is as shown in table 2.As can be seen that traditional PF-MeMBer methods are due to lacking target close to self-adaptive processing, anti-interference energy
Power is poor, target close to or when intersecting can not segmentation object state in time, substantial deviation target time of day, soWithValue it is relatively high；And the method for the present invention with target due to, close to self-adaptive processing mechanism, providing
The anti-interference ability of method, so tracking accuracy will be apparently higher than traditional method.
2 target cross jamming tracking performance of table is evaluated
Experiment 3：Target is newborn and presence is blocked
In this group of experiment scene, image sequence totally 130 frame occurs that target is newborn, disappears and is blocked in different frame
Situations such as, it is tested to verify process performance of the method for the present invention to target new life and disappearance, and to the mesh that is blocked by this group
Target tracking performance.
Figure 10 gives the tracking result of the method for the present invention and tradition PF-MeMBer methods, wherein Figure 10 (a) is tradition
PF-MeMBer method tracking results, Figure 10 (b) are the method for the present invention tracking result., it is apparent that the method for the present invention can
The target for preferably tracking newborn target and being partly blocked, the 122nd, 123,125 and 128 frames as shown in Figure 10 (a) and (b)
In, the method for the present invention can effectively extract newborn target and the state for being at least partially obscured target, and can correctly provide each target
Label, can identify that the movement locus of each target, performance are obviously better than traditional tracking.
Figure 11 is the comparison diagram of the OSPA values of this group experiment, it can be seen that the tracking result of traditional PF-MeMBer methods is missed
Difference is larger, and especially after target 4, which goes out, to be at least partially obscured, the OSPA values of conventional method drastically increase；And the method for the present invention due to
Ability is blocked with preferably anti-, so having higher tracking accuracy, tracking performance is obviously better than conventional method.
This group is tested and carries out 100 Monte Carlo simulations, the performance of tracking accuracy, stability and calculating time etc.
Evaluation is as shown in table 3.As can be seen that traditional PF-MeMBer methods are due to target new life and disappearing insensitive and lack
Target is blocked issue handling, be easy to cause leakage with, with phenomenon, anti-interference ability is poor with mistake, after target is blocked, tracking
Window substantial deviation target time of day, soWithIt is worth relatively high；And the method for the present inventionWithValue all will obviously be less than conventional method, have preferable video multi-target tracking performance.
Table 3 target is newborn and tracking performance is evaluated in the case of being blocked
Claims (6)
1. based on the video multi-target tracking of more Bernoulli Jacob's Eigen Covariances, including：
(1) initialization step：
(1a) initial time k=0, video totalframes are N, and initial target state set isInitialized target stateWhereinIndicate the coordinate in the initial target i rectangle frames upper left corner,Indicate initial target i squares
The width of shape frame,Indicate the height of initial target i rectangle frames, M0Indicate initial target number, initial target existing probability Ps=
0.99, calculate the Eigen Covariance of initial target iParticle initialization sampling；
(1b) initializes newborn dbjective state collectionWherein Indicate newborn
The coordinate in the target i rectangle frames upper left corner,Indicate the width of newborn target i rectangle frames,Indicate the height of newborn target i rectangle frames,
MΓIndicate newborn target number；The newborn hypothesis on location of newborn target is fixed in certain coordinate range, and each frame has target
New life, existing probability 0.02；Calculate the Eigen Covariance of newborn target iThe sampling particle of the newborn target of initialization；
(1c) initialization sampling Fe coatings, particle maximum number are Lmax, particle minimal amount is Lmin, initial target state association
Variance ∑0=diag (1,1,0.1,0.1)；
(2) dbjective state prediction steps：
(2a) assumes that target movement model is random walk model, i.e.,：
X (k+1)=x (k)+e (k)
Wherein, x (k) indicates that k moment dbjective states, e (k) are k moment Gaussian noises, and mean value 0, dbjective state covariance is
(2b) target prediction, it is assumed that can be used more Bernoulli parameter set representations to be in the posterior probability density of k-1 moment multiple targets：
Wherein,WithThe existing probability and probability distribution of k-1 moment targets i, M are indicated respectivelyk-1Indicate k-1 moment target
Number；Posteriority multiple target probability density after then predicting is represented by：
Wherein,Indicate that the k moment survives more Bernoulli parameter collection of target,WithTable respectively
Show that the k moment survives the existing probability of target i and the predicted value of probability distribution,Indicate k moment new life targets
More Bernoulli parameter collection,WithThe existing probability and probability distribution of k moment new life targets i, M are indicated respectivelyΓ,kWhen indicating k
Carve newborn target number；
(3) target likelihood calculates step：
(3a) handles grayscale image sequence and color image sequence, wherein five dimensional features of gray level image extraction are respectively：Ash
Degree, the First-order Gradient in the directions m and n, second order gradient；Coloured image extraction three-dimensional feature be respectively：H, S, V tri- of coloured image
Color value；
The Eigen Covariance T of initialized target template, the Eigen Covariance matrix F of candidate target, and calculate two covariance matrixes
Similarity measure, i.e.,：
D (T, F)=| | log (T)-log (F) | |
Wherein, log () indicates logarithm operation；
(3b) avoids target by partial occlusion, enhances target resolution capability, by target in order to which more accurately target is described
5 piecemeals are expressed as, each block calculates separately its Eigen Covariance matrix, and uses a fusion process, all pieces of fusion
Similarity measure ignores the smallest blocks of similarity measure in candidate target piecemeal, obtains overall similarity and estimates D, i.e.,：
Wherein, min expressions are minimized function,Indicate the Eigen Covariance of the ξ piecemeal of k moment target i templates,
Indicate the Eigen Covariance of the ξ piecemeal of k moment candidate targets i；
After (3c) acquisition overall similarity estimates D, the likelihood of the sampling particle j of k moment candidate targets i is expressed as：
Wherein, [10,30] parameter lambda ∈；
In addition, feature templates TξIt is to be calculated according to the composite character covariance matrix at preceding Φ moment, i.e.,：
Wherein, Fτ,ξIt indicates in τ moment target state estimator statesThe ξ piecemeal Eigen Covariance；
(4) dbjective state updates step：
The k moment, it is assumed that the prediction probability density of video multi-target stochastic finite collection is represented by：
Then updated multiple target posterior probability density is represented by：
Wherein,Indicate the more Bernoulli parameter collection being updated to prediction in the case of missing inspection,Indicate the more Bernoulli parameter collection being updated to measuring collection, ZkIndicate the measurement at k moment
Set, z indicate to measure the element of set；Mk=Mk|k-1+|Zk|, | Zk| indicate the number measured,WithWhen indicating k respectively
Carve the update of the existing probability and probability distribution of target i；
(5) dbjective state extraction step：
In order to avoid sample degeneracy problem, resampling is carried out to sampling particle collection, and estimate according to the sampling updated weights of particle
Count current k moment dbjective state collection Xk, in addition, in order to avoid the redundancy of more Bernoulli parameter collection quantity, there will be probability to be less than
0.01 target is deleted；
(6) target is close to adaptive step：
When two targets close to when, tracking result often by two target discriminations be a target, to cause tracking mistake,
Even there is target leakage with phenomenon；For such case, propose close to self-adaptive processing mechanism；
(7) target following window adaption step：
When video object size varies widely, track window cannot include whole features of target, and information is caused to lose；For
This, proposes track window self-adaptive processing mechanism so that the variation of track window energy adaptive targets size；
(8) particle label step：
By the way that particle is marked, to identify each target, to realize the track following to target.
2. in method for tracking target according to claim 1, wherein step (2b)By one group of weighting particleIt indicates, i.e.,：
Wherein,Indicate the weights of j-th of sampling particle of k-1 moment targets i,Indicate j-th of k-1 moment targets i
The state vector of particle is sampled,Indicate that the sampling particle number of k-1 moment targets i, δ () are Dirac function；
Newborn particle is by newborn target componentDirectly sampling obtains, and the more Bernoulli parameters of survival target indicate
For：
Wherein, pS,kFor k moment target survival probabilities,
3. in method for tracking target according to claim 1, wherein step (4)By one group of weighting particleIt indicates, i.e.,：
Wherein,Indicate the predicted value of the weights of j-th of sampling particle of k moment targets i,Indicate k moment targets i's
The predicted value of the state vector of j-th of sampling particle；
New particle collection carries out more Bernoulli parameter updates,
Wherein, Indicate to measure y to k moment prediction samples particlesLikelihood, this
Place's sampling particle likelihood can be obtained by step (3c).
4. being handled close to objective self-adapting described in method for tracking target according to claim 1, wherein step (6), presses
Following steps carry out：
(6.1) judgement target whether close to
By horizontal direction for, it is assumed that target i and target j, the horizontal distance of two targets are expressed asIfLess than valve
ValueJudgement overlaps on two target level directions, need not handle；And work asMore than threshold valuesAnd it is less than threshold valuesWhen, sentence
Fixed two targets are in the horizontal direction close to and doing separating treatment, the location parameter of more fresh target；
(6.2) close to target separating treatment
wi、wjThe width for indicating the rectangle frame of target i, target j respectively, when judgement two
Close to rear on target level direction, carrying out horizontal separations processing to two targets is：
The state of (6.2.1) target i is xi=[mi,ni,wi,hi], the state of target j is xj=[mj,nj,wj,hj]；
The horizontal distance of (6.2.2) two targets isIfThen arrive step (6.2.3)；Otherwise,
Without carrying out separating treatment；
(6.2.3) is if mi≤mj, then mi=mi-wiα,mj=mj+wjα；Otherwise, mi=mi+wiα,mj=mj-wjα；Wherein, α=
0.15；Vertical direction is identical as horizontal direction method as method for separating and processing close to judging.
5. the target following window adaption processing described in method for tracking target according to claim 1, wherein step (7),
It carries out as follows：
(7.1) target extended mode is obtained
Assuming that in kth frame target i tracking results, the dbjective state extracted isTarget i tracking results
Extended mode be
(7.2) target state estimator state updates
Binary image obtains the minimum rectangle frame for including largest connected region in the rectangle frame of target iThen target state estimator shape
State is updated toParameter beta takes 0.6；
(7.3) treatment on special problems
In some cases, when two targets close to when, the connected region of two targets can be merged into a connected region, or
Due to background interference, a target is made of two or more connected regions, these situations may cause largest connected region to be sent out
Raw mutation；Using the state fusion method at multiple moment, tracking window adaption failure caused by largest connected region mutagenesis is solved
The problem of, i.e.,：
Wherein, setting parameter μ is 0.5, d 5, i.e. the state x of k moment targets ik,iBe by the preceding 5 frame target state with it is current when
It carvesThe value that weighted sum obtains.
6. the particle label technology described in method for tracking target according to claim 1, wherein step (8), by following step
It is rapid to carry out：
(8.1) label prediction, more Bernoulli Jacob's component particle labels at k-1 moment, target of surviving are represented by：
Wherein, Mk-1Indicate the number of Bernoulli Jacob's component,Indicate the population of i-th of Bernoulli Jacob's component, andThe particle label of i.e. same Bernoulli Jacob's component is identical；The particle of newborn more Bernoulli Jacob's components
Label is represented by：
Wherein, MΓIndicate the number of newborn Bernoulli Jacob's component,Indicate the population of i-th of newborn Bernoulli Jacob's component, andThen the prediction of Bernoulli Jacob's component label can be expressed as：Tk|k-1=Tk-1+TΓ；
(8.2) flag update, the label for measuring Bernoulli Jacob's component after updating are represented by：
Wherein, | Zk| it indicates to measure number,
(8.3) resampling, former generation's particle label having the same of particle and they after resampling, the Bernoulli Jacob after resampling
The label of component be：
Wherein MkTo measure Bernoulli Jacob's component number after update,
It can recognize that the movement locus of each target by comparing the label of each Bernoulli Jacob's component；Wherein, when target merges,
The label of the maximum target of existing probability is as the label for merging target.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610860087.9A CN106408594B (en) | 2016-09-28 | 2016-09-28 | Video multi-target tracking based on more Bernoulli Jacob's Eigen Covariances |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610860087.9A CN106408594B (en) | 2016-09-28 | 2016-09-28 | Video multi-target tracking based on more Bernoulli Jacob's Eigen Covariances |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106408594A CN106408594A (en) | 2017-02-15 |
| CN106408594B true CN106408594B (en) | 2018-10-02 |
Family
ID=58015753
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610860087.9A Active CN106408594B (en) | 2016-09-28 | 2016-09-28 | Video multi-target tracking based on more Bernoulli Jacob's Eigen Covariances |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106408594B (en) |
Families Citing this family (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106910205A (en) * | 2017-03-03 | 2017-06-30 | 深圳市唯特视科技有限公司 | A kind of multi-object tracking method based on the coupling of stochastic finite collection wave filter |
| CN106952287A (en) * | 2017-03-27 | 2017-07-14 | 成都航空职业技术学院 | A kind of video multi-target dividing method expressed based on low-rank sparse |
| CN107527356B (en) * | 2017-07-21 | 2020-12-11 | 华南农业大学 | Video tracking method based on lazy interaction mode |
| CN108320302B (en) * | 2018-01-26 | 2019-10-11 | 西安电子科技大学 | CBMeMBer multi-object tracking method based on random hypersurface |
| CN108876819A (en) * | 2018-06-11 | 2018-11-23 | 深圳市唯特视科技有限公司 | A kind of three-dimensional multi-target tracking algorithm based on the more Bernoulli Jacob's filtering of Poisson |
| CN109297478B (en) * | 2018-09-19 | 2022-05-17 | 西安汇智信息科技有限公司 | Fiber-optic gyroscope navigation self-adaptive filtering method based on GM-CBMeMBer |
| CN109508444B (en) * | 2018-12-18 | 2022-11-04 | 桂林电子科技大学 | Quick tracking method for interactive multimode generalized label multi-Bernoulli under interval measurement |
| CN109656271B (en) * | 2018-12-27 | 2021-11-02 | 杭州电子科技大学 | Track soft association method based on data association idea |
| CN110598830B (en) * | 2019-04-03 | 2021-05-11 | 常熟理工学院 | Joint multi-cell tracking method based on label ant colony |
| CN110084831B (en) * | 2019-04-23 | 2021-08-24 | 江南大学 | Multi-target detection tracking method based on YOLOv3 multi-Bernoulli video |
| US11094070B2 (en) * | 2019-04-23 | 2021-08-17 | Jiangnan University | Visual multi-object tracking based on multi-Bernoulli filter with YOLOv3 detection |
| CN110967690B (en) * | 2019-11-12 | 2023-06-16 | 江南大学 | Multi-target tracking method based on multiple Bernoulli distributed multiple sensors |
| CN111311006B (en) * | 2020-02-21 | 2022-08-16 | 河海大学常州校区 | State prediction and production energy-saving scheduling method for reentry Bernoulli production system |
| CN111488552B (en) * | 2020-04-24 | 2023-03-21 | 商丘师范学院 | Close-proximity multi-target tracking method based on Gaussian mixture probability hypothesis density |
| CN111754545B (en) * | 2020-06-16 | 2024-05-03 | 江南大学 | IOU (input-output unit) matching-based double-filter video multi-target tracking method |
| CN111964706B (en) * | 2020-08-13 | 2022-07-05 | 桂林电子科技大学 | Distributed field-of-view complementary multi-Bernoulli correlation arithmetic mean fusion tracking method |
| CN111739024B (en) * | 2020-08-28 | 2020-11-24 | 安翰科技(武汉)股份有限公司 | Image recognition method, electronic device and readable storage medium |
| CN112305915B (en) * | 2020-10-28 | 2023-02-07 | 深圳大学 | Label multi-Bernoulli multi-target tracking method and system of TSK iterative regression model |
| CN114061584B (en) * | 2021-11-02 | 2024-05-24 | 江苏科技大学 | Multi-robot-based potential equalization Bernoulli filtering SLAM method |
| CN114690175B (en) * | 2022-04-12 | 2023-10-27 | 电子科技大学 | Target direct detection and tracking method based on passive external radiation source radar |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6541689B1 (en) * | 1999-02-02 | 2003-04-01 | Microsoft Corporation | Inter-track communication of musical performance data |
| CN103176164A (en) * | 2013-04-11 | 2013-06-26 | 北京空间飞行器总体设计部 | Multi-target passive tracking method based on wireless sensor network |
| CN103345577A (en) * | 2013-06-27 | 2013-10-09 | 江南大学 | Probability hypothesis density multi-target tracking method based on variational Bayesian approximation technology |
| CN103678949A (en) * | 2014-01-09 | 2014-03-26 | 江南大学 | Tracking measurement set partitioning method for multiple extended targets based on density analysis and spectrum clustering |
| CN104766320A (en) * | 2015-04-02 | 2015-07-08 | 西安电子科技大学 | Bernoulli smoothing weak target detection and tracking method under thresholding measuring |
| CN104777465A (en) * | 2014-01-09 | 2015-07-15 | 江南大学 | Random extended object shape and state estimation method based on B spline function |
| CN105354860A (en) * | 2015-08-26 | 2016-02-24 | 西安电子科技大学 | Box particle filtering based extension target CBMeMBer tracking method |
-
2016
- 2016-09-28 CN CN201610860087.9A patent/CN106408594B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6541689B1 (en) * | 1999-02-02 | 2003-04-01 | Microsoft Corporation | Inter-track communication of musical performance data |
| CN103176164A (en) * | 2013-04-11 | 2013-06-26 | 北京空间飞行器总体设计部 | Multi-target passive tracking method based on wireless sensor network |
| CN103345577A (en) * | 2013-06-27 | 2013-10-09 | 江南大学 | Probability hypothesis density multi-target tracking method based on variational Bayesian approximation technology |
| CN103678949A (en) * | 2014-01-09 | 2014-03-26 | 江南大学 | Tracking measurement set partitioning method for multiple extended targets based on density analysis and spectrum clustering |
| CN104777465A (en) * | 2014-01-09 | 2015-07-15 | 江南大学 | Random extended object shape and state estimation method based on B spline function |
| CN104766320A (en) * | 2015-04-02 | 2015-07-08 | 西安电子科技大学 | Bernoulli smoothing weak target detection and tracking method under thresholding measuring |
| CN105354860A (en) * | 2015-08-26 | 2016-02-24 | 西安电子科技大学 | Box particle filtering based extension target CBMeMBer tracking method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106408594A (en) | 2017-02-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106408594B (en) | Video multi-target tracking based on more Bernoulli Jacob's Eigen Covariances | |
| CN105405151B (en) | Anti-shelter target tracking based on particle filter and weighting Surf | |
| CN110084831A (en) | Based on the more Bernoulli Jacob's video multi-target detecting and tracking methods of YOLOv3 | |
| CN103729854B (en) | A kind of method for detecting infrared puniness target based on tensor model | |
| Zhang et al. | Multiple object tracking via species-based particle swarm optimization | |
| CN103971386B (en) | A kind of foreground detection method under dynamic background scene | |
| CN111932583A (en) | Space-time information integrated intelligent tracking method based on complex background | |
| CN110147807A (en) | A kind of ship intelligent recognition tracking | |
| CN105046195B (en) | Human bodys' response method based on asymmetric generalized gaussian model | |
| CN109800689A (en) | A kind of method for tracking target based on space-time characteristic fusion study | |
| CN110298865B (en) | Space-based starry sky background weak small point target tracking method based on threshold separation clustering device | |
| CN110298297A (en) | Flame identification method and device | |
| CN109191488B (en) | Target tracking system and method based on CSK and TLD fusion algorithm | |
| CN108460790A (en) | A kind of visual tracking method based on consistency fallout predictor model | |
| CN106127812A (en) | A kind of passenger flow statistical method of non-gate area, passenger station based on video monitoring | |
| CN109033978A (en) | A kind of CNN-SVM mixed model gesture identification method based on error correction strategies | |
| CN106952293A (en) | A kind of method for tracking target based on nonparametric on-line talking | |
| CN102289822A (en) | Method for tracking moving target collaboratively by multiple cameras | |
| CN107622507A (en) | A kind of air target tracking method based on deep learning | |
| CN106127798B (en) | Dense space-time contextual target tracking based on adaptive model | |
| CN105321188A (en) | Foreground probability based target tracking method | |
| CN112233145A (en) | Multi-target shielding tracking method based on RGB-D space-time context model | |
| Liu et al. | [Retracted] Self‐Correction Ship Tracking and Counting with Variable Time Window Based on YOLOv3 | |
| CN114038011A (en) | Method for detecting abnormal behaviors of human body in indoor scene | |
| CN116342645A (en) | Multi-target tracking method for natatorium scene |
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 |