CN109636834A - Video frequency vehicle target tracking algorism based on TLD innovatory algorithm - Google Patents
Video frequency vehicle target tracking algorism based on TLD innovatory algorithm Download PDFInfo
- Publication number
- CN109636834A CN109636834A CN201811395493.8A CN201811395493A CN109636834A CN 109636834 A CN109636834 A CN 109636834A CN 201811395493 A CN201811395493 A CN 201811395493A CN 109636834 A CN109636834 A CN 109636834A
- Authority
- CN
- China
- Prior art keywords
- target
- particle
- frame
- classifier
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/269—Analysis of motion using gradient-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/277—Analysis of motion involving stochastic approaches, e.g. using Kalman filters
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
Abstract
Video frequency vehicle target tracking algorism based on TLD innovatory algorithm, first input video, calibration tracking target;Secondly target is tracked using LK optical flow tracking device, by the way that by the uniform gridding of video frame, the top left corner apex for choosing each grid is characterized a little, then comes the position of tracking prediction these characteristic points in the next frame using forward-backward algorithm track optical flow method;And optimization is improved to TLD tracker module, introduces Kalman filter and particle filter, while being scanned with random fern classifier, generates a binary coding x.Encode the posterior probability P that x is directed toward some leaf node of decision treei(y x), wherein (0,1) y ∈.Then the average value of the posterior probability of all mutually independent decision tree outputs is acquired;Then learn to update classifier by P-N;Finally by comprehensive assessment, the highest tracking target of accuracy is shown.
Description
Technical field
Present invention is primarily based on the researchs to single body long-time track algorithm in video, and are applied to vehicle inspection
Survey and target tracking domain, and from the angle of practical application, it proposes a kind of track algorithm based on machine learning, belongs to mesh
Mark tracing detection related fields.
Background technique
Vision is one of the important channel in the human cognitive world, in mankind's external world information obtained about from
Human visual system, human visual system assume responsibility for the processing work of the bulk information in human lives, it can be fast and accurately
Complete the tasks such as pattern imaging, description, identification and understanding.Computer vision be exactly understand human vision it is essential on the basis of,
Human sight apparatus is replaced with various imaging systems, replaces human brain to complete processing and explanation to input picture with computer.Meter
The final goal in research of calculation machine vision is exactly to enable a computer to pass through visual observation as the mankind, understand the world, and final
With the autonomous ability for adapting to environment.
Hot subject one of of the computer vision as artificial intelligence field, it is a comprehensive subject, is attracted
Researcher from every subjects is added among the research to it, it has merged signal processing, computer science and engineering object
The research method and achievement of the ambits such as of science, applied mathematics and statistics, neuro-physiology.
Likewise, video frequency object tracking, which is also one, has merged multi-disciplinary complicated project.Including image procossing, mould
Formula identification and random process and probability theory and partial differential equation etc..According to the specific mistake for realizing tracking in Target Tracking System
Journey has summed up general method for tracking target architecture from a kind of completely new angle.Its method structure is divided into following three
Partially (1) tracking clarification of objective is chosen and is indicated;(2) the common algorithm frame of tracking of target is tracked;(3) tracking target is pre-
Method of determining and calculating.The detailed process of current understood target following is divided into following several steps, and first step needs detect in depending on video sequence
Exist out effective target region or effective target;Second step carries out the segmentation of science to the effective target that detected;
The feature of effective target is extracted, and forms effective object matching information model;It is predicted according to prediction model in lower a period of time
The location information that target is likely to occur is carved, to lock effective search range;Third step is before using in the search range of prediction
The target information template at one moment carries out effective target matching, to find optimal matching position.If what is predicted in advance
Effective target is not found in range, it is necessary to carry out specific scientific disposal.Method is first with the suspected target being matched to
Correct the information model of effective target tracking.And repeat above-mentioned three-step process.
Hot research topic one of of the video frequency object tracking technology as computer vision field, receives domestic and foreign scholars
With the extensive concern of research institution.So-called video frequency object tracking technology, which refers to, imitates human visual system by computer mould, passes through
Analysis to image sequence obtained by camera, calculates the location parameter of user's interesting target, such as the two-dimensional coordinate position of target
It sets, the size of image-region shared by target, target etc., and according to the different characteristic of target, to same in image sequence
Moving target is associated, and obtains the entire motion track of the moving target.By the development of many years, video frequency object tracking technology
The many aspects being widely used in life and military affairs.
The moving vehicles detection and tracking studied herein is wherein important application.With the development of economy, vehicle is gradually
Increase, stored digital, computing capability and the video compression standard of rapid development lead to the strong growth of video content, produce sea
The road video data of amount, personal monitoring are a uninteresting and time-consuming job.The video detection technology process of view-based access control model
The adverse effect of few human factors (absent minded, respond is slow etc.), and can with 24 hours one day not between
Disconnected work, will not report the abnormal conditions occurred in video image frame by mistake whiles saving a large amount of human and material resources, financial resources etc.
With fail to report.This makes the automatic monitoring technical based on video have important value, significant.
Summary of the invention
The purpose of the present invention is intended to improve the accuracy of long-time video frequency object tracking, and improves the anti-of track algorithm and block
Ability.
In order to achieve the above objectives, the present invention proposes a kind of video frequency vehicle target tracking algorism based on TLD innovatory algorithm, packet
Include following steps:
Step 1, input video, calibration tracking target;
Step 2.1, target is tracked using LK optical flow tracking device, and proposes that tracker improves and optimizates method;
Step 2.2, it is scanned with random fern classifier;
Step 3, learn to update classifier by P-N;
Step 4, display tracking target.
Video frequency vehicle target tracking algorism based on TLD innovatory algorithm, which comprises the following steps:
The first step, input video, calibration tracking target;
Second step initializes LK optical flow tracking device and random fern classifier, is scanned and tracks to target;
Tracker in TLD uses a kind of LK optical flow method based on forward-backward algorithm track;Present frame is It next frame
For It+1, light stream rule in forward-backward algorithm track is after predicting It+1 by It, using the point predicted in It+1, then does anti-
To prediction, i.e., It is predicted by It+1, obtains an offset deviation by this forward-backward algorithm trajectory predictions;If backward prediction
The obtained key point displacement deviation in the characteristic point and original known It in It is greater than threshold value 16, then will predict in It+1
The biggish characteristic point of deviation exclude;In having cast out present frame after the biggish characteristic point of offset deviation, just obtain current
Complete corresponding point in frame and next frame;Template is done to the image-region around mutual corresponding point in two frame of front and back respectively
Match, calculates the similarity between image-region, once similarity is less than the intermediate value of all image-region similarities, then by these
The small future position of similarity further excludes;The median in the direction x and the direction y offset between remaining corresponding points is calculated separately,
The dimensional variation factor as new prediction block in the direction x and y, the position of next frame prediction block is found out further according to the dimensional variation factor
It sets and size;And so on, obtain preliminary tracking result;
Random fern classifier is made of many a fundamental classifiers;The process object of classifier and the process object of tracker
It is identical, it is current image frame, and the work of classifier and detector carries out simultaneously;I is in image block for each fundamental classifier
On according to it is initial when the pixel that determines to acquisition pixel to the difference of gray scale, generate a binary coding x;Encode x
It is directed toward the posterior probability P of some leaf node of decision treei(y x), wherein (0,1) y ∈;Then acquire it is all it is mutually independent certainly
The average value of the posterior probability of plan tree output;Image block of the average value greater than 50% is by the classifier, and output result is as mark
Remember that sample enters next module;
Third step learns to update classifier by P-N;
By priori signature sample and unlabelled sample come Study strategies and methods;Marker samples derive from the output of classifier
As a result;Study is made of the constraint of two class formations, i.e., just constrain and break a promise beam;It constrains and not labeled sample is marked point
Class trains classifier later;Positive constraint refers to the constraint condition that unknown sample is labeled as to positive sample, and it is attached to will be close to track here
Close sample labeling is positive sample;Beam of breaking a promise refers to the constraint condition that unknown sample is labeled as to negative sample, will be far from rail here
The sample labeling of mark is negative sample;
If x is characterized a sample in space X, y indicates a label in corresponding label space Y={ -1,1 },
So sample space and corresponding label are indicated with set { X, Y };P-N learns according to marked sample set { Xl, Yl }
Classifier, and training sample are established, guides classifier to work using not marked data Xu;
4th step, display tracking target;
It is updated according to P-N study mechanism with note fern classifier, shows tracking result, the target for recycling classifier to judge
The target frame that frame and tracker predict, is compared with realistic objective, and comprehensive descision goes out final accurately track as a result, with mesh
The mode of mark frame is shown in video.
Tracker module is improved using Kalman filter or particle filter;
S1:Kalman filters improved method
State vector of the target at the k moment indicates are as follows: Xk=[xk, yk, x 'k, y 'k]T, wherein xk, yKIt is illustrated respectively in x, y
Coordinate on direction, xk, yKSpeed of the target on the direction x, y is respectively indicated, k-1 indicates last moment;
The position of target is chosen as observation vector, observation vector indicates are as follows: z (k)=[xck, yck]T, wherein, xck、yck
Respectively indicate the position coordinates by observing target's center obtained on the direction x, y;
What the center of target was done is to become to accelerate linear motion, acceleration wk-1Random variation, and Gaussian distributed is
wk-1~N (0, σ2 w);
According to Newton's laws of motion:
xk=xk-1+x′k-1t+0.5wk-1t2
yk=yk-1+y′k-1t+0.5wk-1t2
x′k=x 'k-1+wk-1t
y′k=y 'k-1+wk-1t
So thus obtain process model;According to Xk=AkXk-1+Cwwk-1, it obtains:
By observation model Zk=HkXk+CvVkIt obtains:
Frame per second is indicated using t, then state-transition matrix, observing matrix indicate are as follows:
The state of initial time target is set:Wherein, x0·y0Indicate that target is in the direction x in first frame
With the position in the direction y;0,0 initial velocity for respectively indicating target target on the direction x, y;
Initial observed quantity Z is set0=[x0, y0]T;
It is Q that initial state covariance, which is arranged,k, i.e. plant noise, state measurement error co-variance matrix Pk, state transfer
Matrix Ak, initial observation covariance Rk, i.e. observation noise, observing matrix Hk;Specifically it is provided that
After setting initial parameter, the recurrence of Kalman filtering is carried out using following steps:
(1) state at current time is predicted, includes speed and position;
(2) prior estimate error covariance is calculated;
(3) observation and status predication value are utilized, the optimal value at current time is obtained;
(4) kalman gain is calculated;
(5) error covariance is updated;
S2: particle filter improved method
(1) the state selection of particle
Utilize Xk=(x, y, s) indicates the state of particle, wherein (x, y) indicates the coordinate position of particle in the video frame,
Namely rectangle frame center corresponding to particle, s indicate the dimensional variation factor of rectangle frame;200 particles are chosen, by 200
The position initialization of a particle is the center of initial target frame, i.e. X0=(x0, y0, s0), scale is initialized as 1, and
Calculate hsv color histogram corresponding to initial target frame;
(2) the state transfer of system
Particle moves to the position in next from the position in previous frame, needs to use state transfer, common state
There are two types of transfer methods;Random transferring and second-order auto-regressive transfer;
Random transferring is exactly the center position random distribution particle in previous frame;;
Second-order auto-regressive metastasis model predicts particle in next frame using the random combine of the particle state of previous instant
In position, common second-order autoregressive model is expressed as follows:
Wherein, XkIndicate the state of k moment particle,Indicate the mean value of all particle estimations, Xk-1Indicate k-1 moment particle
State, Xk-2Indicate the state of k-2 moment particle, w indicates random noise;A1、A2, B be all constant, B indicates the propagation of particle
Radius;It is obtained according to second-order autoregressive model, status predication of the particle at the k moment:
These parameters it is known that
A1=diag { a1, a2, a3, A2=diag { a2, a2, a2, B=diag { b, b, b },
a1=2.0, a2=1.0, a3=1.0, b=1.0, wk∈ (0,0.001);
If the mean value of all particle estimations is (x0, y0, 1.0), wherein ((x0, y0) indicate previous frame in target centre bit
It sets, 1.0 indicate average dimension;Then the propagation of single particle calculates as follows:
xk=a1(xk-1-x0)+a2(xk-2-x0)+bwk+x0
yk=a1(yk-1-x0)+a2(yk-2-x0)+bwk+y0
sk=a1(sk-1-x0)+a2(sk-2-x0)+bwk+1.0
(3) observation model is established
It after predicting the position of particle in the next frame, needs to be observed particle, that is, judges each particle institute
Similarity degree between the target and true target state of representative;;
IfIndicate HSV field color histogram corresponding to candidate particle x,It indicates corresponding to reference target model
HSV field color histogram, then the Bhattacharyya coefficient between two color histograms is expressed as:
Wherein, u indicates the dimension of vector,Indicate the vector of m dimension;;
Define the measurement distribution of color histogram:
(4) importance sampling
The weight of more new particle, using target frame corresponding to the particle of maximum weight as final target frame position;
(5) resampling
According to the weight of each particle, descending sequence is carried out, calculates its average value;In the certain feelings of total number of particles
Under condition, the number of particles that weight is greater than average value is increased, weight is less than number corresponding to the particle of average value and is reduced, very
To give up weight be less than average value particle.
Beneficial effect
Although TLD algorithm has good re-detection ability and for a long time ability of study, block in reply, shape
When the influences of factors such as change, the tracking effect of the algorithm is ideal not enough.Therefore Kalman and particle filter are introduced in tracking module
Wave, Kalman filter have good predictive estimation ability, and particle filter is then very suitable to processing nonlinear motion problem.It will
TLD is combined with the two respectively, and a kind of improved on-line study mechanism is added, so that whole target following is more
Stablize, effectively.TLD algorithm is to determine the positions and dimensions of simple target in initial image frame.Then next each
Picture frame, algorithm will test the positions and dimensions of target or indicate that target whether there is, then track frame by frame, detector detection
The position that target is occurred, and tracker is corrected when needed.The mistake of P-N learning evaluation detector, and update inspection
Device is surveyed to improve the performance of detector.
Detailed description of the invention
Above-mentioned and/or additional aspect and advantage of the invention will become from the following description of the accompanying drawings of embodiments
Obviously and it is readily appreciated that, in which:
Fig. 1 is the flow chart of the video frequency vehicle target tracking algorism based on TLD algorithm of the embodiment of the present invention;
Fig. 2 is the kalman innovatory algorithm principle flow chart of one embodiment of the invention.
Fig. 3 is the particle filter innovatory algorithm principle flow chart of one embodiment of the invention.
Fig. 4 is the random fern classifier work of the video frequency vehicle target tracking algorism based on TLD algorithm of the embodiment of the present invention
Make schematic diagram;
Fig. 5 is the tracking result of the video frequency vehicle target tracking algorism based on TLD algorithm of the embodiment of the present invention.
Specific embodiment
The embodiment of the present invention is described below in detail, examples of the embodiments are shown in the accompanying drawings, wherein from beginning to end
Same or similar label indicates same or similar element or element with the same or similar functions.Below with reference to attached
The embodiment of figure description is exemplary, and for explaining only the invention, and is not construed as limiting the claims.
As shown in Figure 1, the present invention proposes a kind of video frequency vehicle target tracking algorism based on TLD innovatory algorithm, it is specific to walk
It is rapid as follows:
Step 1, input video, calibration tracking target;
Step 2.1, target is tracked using LK optical flow tracking device, and proposes that tracker improves and optimizates method;
So-called light stream refers to the speed that grayscale mode moves in image.It is that the three dimensional velocity vectors of visible point in scenery exist
Projection on imaging plane, it illustrates the instantaneous variation of the point on scenery surface position in the picture.
Three of LK method hypothesis: (1) with the variation of interframe movement, brightness do not occur for brightness constancy, i.e. same point
Change.This is the hypothesis of basic optical flow method;(2) Time Continuous, i.e. image change with time very slowly;(3) space is consistent,
It is also neighbor point on image that neighbouring point, which projects to, in one scene, and neighbouring spot speed is consistent.This is the distinctive vacation of optical flow method
It is fixed.
(1) assume that original image is I (x, y, z) (three-dimensional space being expanded to here, so there are also a z values), after mobile
Image is I (x+ δ x, y+ δ y, z+ δ z, t+ δ t), and the two meets:
t+H.O.T.
(H.O.T. refers to higher order, can ignore in the case where movement is sufficiently small)
(2) wherein image is mobile can consider I (x, y, z, t)=I (x+ δ x, y+ δ y, z+ δ z, t+ δ t) that is:
(3) we are available from this equation:Wherein Vx=u, Vy=v,
The namely value (two dimensional image does not have z) of light stream,Then image (x, y, z, t) this point gradient (It is two frames
Difference between image block).
(4) assume that stream (Vx, Vy, Vz) is a constant in the small window that a size is m*m*m (m > 1), then from picture
Available following one group of equation in plain 1...n, n=m*m*m:
Ix1Vx+Iy1Vy+Iz1Vz=-It1;Ix2Vx+Iy2Vy+Iz2Vz=-It2;IxnVx+IynVy+IznVz=-Itn
Three unknown numbers still have more than three equations, this equation group is an overdetermined equation naturally, that is to say, that side
There is redundancy in journey group, equation group can indicate are as follows:
NamelyIn order to solve this problem, using least square method:
It obtains:
Wherein summation is from 1 to n.
Using Kalman filter and the good prediction characteristic of particle filter, tracker module is improved.Two kinds
Improved method be it is arranged side by side, preliminary tracking result is impacted without ordinal relation, and only, does not influence the rear mold of classifier sum
The work of block.
Kalman filter and algorithm improvement method: although TLD track algorithm has good tracking performance, TLD exists
Tracking effect is unobvious when encountering target occlusion, or even the case where tracking error and tracking loss occurs.Kalman filter is
It is proved to have good predictive ability, therefore Kalman filter and TLD track algorithm is combined, is proposed based on Kalman
The TLD track algorithm of filtering, tracking ability of the Lai Tigao track algorithm under target occlusion.
After input video frame, the dbjective state in video frame is input to Kalman filter, then according to Kalman
The predictive equation of filter predicts target in the state at moment once, meanwhile, using TLD algorithm also to the subsequent time of target
State tracked, determine the tracing area of target.When TLD tracking effect is relatively good, using the tracking result of TLD as
Observation is adjusted update to the parameter of Kalman filter.When TLD causes target to be lost due to blocking, then utilize
Kalman filter predicts the position of target, and the update of Kalman filter is carried out using the predicted value as observation, from
And it improves the anti-of algorithm and blocks tracking ability.Principle is as shown in Figure 2.
Particle filter and algorithm improvement method: Kalman filter is but big in reality system for solving the problems, such as Gauss
It is not linear existing for amount or Gaussian Profile, but non-linear, even Complete heart block sometimes.Such as in pattern-recognition
Field, the extraction of feature are all based on two-dimensional image matrix, this is Complete heart block, for describing the face of image color information
Color Histogram feature is also non-linear.In this case, it is practically impossible to obtain Chapman-Kolmogoroff integration type
Analytic solutions, Bayesian filter frame can not practical application.For this purpose, monte carlo method was suggested in 1940, base therewith
In the frame that the sequential importance sampling SIS method of Monte Carlo thought is introduced into Bayesian filter.
Importance sampling SIS method is introduced for solving the problems, such as in Bayesian filter that nonlinear and non-Gaussian can not parse
The problem of calculating.Particle filter indicates the prior probability function in Bayesian filter using a large amount of sampled points, thus by PDF
Function expansion has arrived the arbitrary form of non-gaussian, obtains the particle of weighted using SIS, adjustable by the difference of weight
The distribution of particle, so as to solve the problems, such as non-gaussian.
After video frame input, using the training set of TLD algorithm, three steps of classifier and classification results are sentenced
It is disconnected, judge preliminary tracking prediction frame.Meanwhile starting particle filter algorithm tracking.With common particle filter tracking algorithm
Difference then reinitializes particle filter once the confidence level of the tracking prediction frame of TLD algorithm is more than threshold value, and by particle
The output of filtering is as final tracking prediction frame.Principle is as shown in Figure 3.
Step 2.2, it is scanned with random fern classifier;
The process object of classifier and the process object of tracker are identical, are current image frame, and classifier and detection
The work of device carries out simultaneously.Classified using random fern jungle the part.Random fern jungle is one and includes multiple random ferns
Classifier.The pixel of each random fern k initial target and candidate region acquisition pixel gray scale is made the difference,
Generate corresponding binary coding X.The posterior probability of X direction decision tree.These posterior probability of all decision trees are taken
Semi-naive Bayes output valve M, then the precondition that can enter following nearest neighbor classifier is, M is greater than 50%.When this kind
It is judged that current image block contains target when happening, reject region is otherwise entered.Its principle is as shown in Figure 4.
Step 3, learn to update classifier by P-N;
The machine learning method that TLD is used is P-N study.P-N study is a kind of semi-supervised machine learning algorithm, its needle
The two kinds of mistakes generated when to classifier to sample classification provide two kinds " experts " and correct: P expert: detecting missing inspection just
Sample;N expert: the positive sample of erroneous detection is corrected.
Image is progressively scanned with various sizes of scanning window, often just forms an encirclement frame a position, is wrapped
As soon as image-region determined by peripheral frame is known as an image primitive (patch), the sample set that image primitive enters machine learning becomes one
A sample.The sample that scanning generates is non-exemplar, needs to be classified with classifier, determines its label.
If algorithm have determined object in the position (the corresponding position for surrounding frame has actually been determined) of t+1 frame, from
Filtered out in the encirclements frame that detector generates 10 and it apart from nearest encirclement frame (area of two friendships for surrounding frame is divided by simultaneously
Area be greater than 0.7), to each encirclements frame do small affine transformation (translation 10%, scaling 10%, rotation 10 ° within), production
Raw 20 image primitives, thus generate 200 positive samples.Select that several (area of friendship is divided by simultaneously apart from farther away encirclement frame again
Area less than 0.2), generate negative sample.The sample generated in this way is the sample of label, these samples are put into training set,
For updating the parameter of classifier.
Step 4, display tracking target, as shown in Figure 5.
Claims (2)
1. the video frequency vehicle target tracking algorism based on TLD innovatory algorithm, which comprises the following steps:
The first step, input video, calibration tracking target;
Second step initializes LK optical flow tracking device and random fern classifier, is scanned and tracks to target;
Tracker in TLD uses a kind of LK optical flow method based on forward-backward algorithm track;Present frame is that It next frame is It+
1, light stream rule in forward-backward algorithm track is after predicting It+1 by It, using the point predicted in It+1, then does reversed pre-
It surveys, i.e., It is predicted by It+1, obtain an offset deviation by this forward-backward algorithm trajectory predictions;If backward prediction obtains
It in characteristic point and original known It in key point displacement deviation be greater than threshold value 16, then it is inclined by what is predicted in It+1
The biggish characteristic point of difference excludes;In having cast out present frame after the biggish characteristic point of offset deviation, just obtain present frame and
Complete corresponding point in next frame;Template matching is done to the image-region around mutual corresponding point in two frame of front and back respectively, is counted
The similarity between image-region is calculated, once similarity is less than the intermediate value of all image-region similarities, then and it is these are similar
Small future position is spent further to exclude;The median for calculating separately the direction x and the direction y offset between remaining corresponding points, as
New prediction block the direction x and y the dimensional variation factor, further according to the dimensional variation factor find out next frame prediction block position and
Size;And so on, obtain preliminary tracking result;
Random fern classifier is made of many a fundamental classifiers;The process object phase of the process object of classifier and tracker
Together, it is current image frame, and the work of classifier and detector carries out simultaneously;I is on image block for each fundamental classifier
According to it is initial when the pixel that determines to acquisition pixel to the difference of gray scale, generate a binary coding x;Coding x refers to
To the posterior probability P of some leaf node of decision treei(y x), wherein (0,1) y ∈;Then all mutually independent decisions are acquired
Set the average value of the posterior probability of output;Image block of the average value greater than 50% is by the classifier, and output result is as label
Sample enters next module;
Third step learns to update classifier by P-N;
By priori signature sample and unlabelled sample come Study strategies and methods;Marker samples derive from the output knot of classifier
Fruit;Study is made of the constraint of two class formations, i.e., just constrain and break a promise beam;It constrains and classification is marked to not labeled sample,
Classifier is trained later;Positive constraint refers to the constraint condition that unknown sample is labeled as to positive sample, will be close near track here
Sample labeling be positive sample;Beam of breaking a promise refers to the constraint condition that unknown sample is labeled as to negative sample, will be far from track here
Sample labeling be negative sample;
If x is characterized a sample in space X, y indicates a label in corresponding label space Y={ -1,1 }, then
Sample space and corresponding label are indicated with set { X, Y };P-N study is built according to marked sample set { Xl, Yl }
Vertical classifier, and training sample, guide classifier to work using not marked data Xu;
4th step, display tracking target;
Updated according to P-N study mechanism with note fern classifier, show tracking result, recycle the target frame judged of classifier and
The target frame that tracker predicts, is compared with realistic objective, and comprehensive descision goes out final accurately track as a result, with target frame
Mode show in video.
2. according to the method described in claim 1, it is characterized by: using Kalman filter or particle filter to tracker
Module improves;
S1:Kalman filters improved method
State vector of the target at the k moment indicates are as follows: Xk=[xk, yk, x 'k, y 'k]T, wherein xk、ykIt is illustrated respectively in the direction x, y
On coordinate, xk、ykSpeed of the target on the direction x, y is respectively indicated, k-1 indicates last moment;
The position of target is chosen as observation vector, observation vector indicates are as follows: z (k)=[xck, yck]T, wherein, xck、yckRespectively
Indicate the position coordinates by observing target's center obtained on the direction x, y;
What the center of target was done is to become to accelerate linear motion, acceleration wk-1Random variation, and Gaussian distributed, that is, wk-1
~N (0, σ2 w);
According to Newton's laws of motion:
xk=xk-1+x’k-1t+0.5wk-1t2
yk=yk-1+y′k-1t+0.5wk-1t2
x′k=x 'k-1+wk-1t
y′k=y 'k-1+wk-1t
So thus obtain process model;According to Xk=AkXk-1+Cwwk-1, it obtains:
By observation model Zk=HkXk+CvVkIt obtains:
Frame per second is indicated using t, then state-transition matrix, observing matrix indicate are as follows:
The state of initial time target is set:Wherein, x0、y0Indicate that target is in the direction x and the side y in first frame
To position;0,0 initial velocity for respectively indicating target target on the direction x, y;
Initial observed quantity Z is set0=[x0, y0]T;
It is Q that initial state covariance, which is arranged,k, i.e. plant noise, state measurement error co-variance matrix PkState-transition matrix
Ak, initial observation covariance Rk, i.e. observation noise, observing matrix Hk;Specifically it is provided that
After setting initial parameter, the recurrence of Kalman filtering is carried out using following steps:
(1) state at current time is predicted, includes speed and position;
(2) prior estimate error covariance is calculated;
(3) observation and status predication value are utilized, the optimal value at current time is obtained;
(4) kalman gain is calculated;
(5) error covariance is updated;
S2: particle filter improved method
(1) the state selection of particle
Utilize Xk=(x, y, s) indicates the state of particle, wherein (x, y) indicates the coordinate position namely grain of particle in the video frame
Rectangle frame center corresponding to son, s indicate the dimensional variation factor of rectangle frame;200 particles are chosen, by 200 particles
Position initialization be initial target frame center, i.e. X0=(x0, y0, s0), scale is initialized as 1, and calculate just
Hsv color histogram corresponding to beginning target frame;
(2) the state transfer of system
Particle moves to the position in next from the position in previous frame, needs to use state transfer, common state transfer
There are two types of methods;Random transferring and second-order auto-regressive transfer;
Random transferring is exactly the center position random distribution particle in previous frame;;
Second-order auto-regressive metastasis model predicts using the random combine of the particle state of previous instant particle in the next frame
Position, common second-order autoregressive model are expressed as follows:
Wherein, XkIndicate the state of k moment particle,Indicate the mean value of all particle estimations, Xk-1Indicate the shape of k-1 moment particle
State, Xk-2Indicate the state of k-2 moment particle, w indicates random noise;A1、A2, B be all constant, B indicates the propagation radius of particle;
It is obtained according to second-order autoregressive model, status predication of the particle at the k moment:
These parameters it is known that
A1=diag { a1, a2, a3, A2=diag { a2, a2, a2, B=diag { b, b, b }, a1=2.0, a2=1.0, a3=1.0,
B=1.0, wk∈ (0,0.001);
If the mean value of all particle estimations is (x0, y0, 1.0), wherein (x0, y0) indicate previous frame in target center, 1.0
Indicate average dimension;Then the propagation of single particle calculates as follows:
xk=a1(xk-1-x0)+a2(xk-2-x0)+bwk+x0
yk=a1(yk-1-x0)+a2(yk-2-x0)+bwk+y0
sk=a1(sk-1-x0)+a2(sk-2-x0)+bwk+1.0
(3) observation model is established
It after predicting the position of particle in the next frame, needs to be observed particle, that is, judges representated by each particle
Target and true target state between similarity degree;;
IfIndicate HSV field color histogram corresponding to candidate particle x,Indicate the corresponding HSV of reference target model
Field color histogram, then the Bhattacharyya coefficient between two color histograms is expressed as:
Wherein, u indicates the dimension of vector,Indicate the vector of m dimension;;
Define the measurement distribution of color histogram:
(4) importance sampling
The weight of more new particle, using target frame corresponding to the particle of maximum weight as final target frame position;
(5) resampling
According to the weight of each particle, descending sequence is carried out, calculates its average value;In the certain situation of total number of particles
Under, the number of particles that weight is greater than average value is increased, weight is less than number corresponding to the particle of average value and is reduced, even
Give up the particle that weight is less than average value.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811395493.8A CN109636834A (en) | 2018-11-22 | 2018-11-22 | Video frequency vehicle target tracking algorism based on TLD innovatory algorithm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811395493.8A CN109636834A (en) | 2018-11-22 | 2018-11-22 | Video frequency vehicle target tracking algorism based on TLD innovatory algorithm |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109636834A true CN109636834A (en) | 2019-04-16 |
Family
ID=66068753
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811395493.8A Pending CN109636834A (en) | 2018-11-22 | 2018-11-22 | Video frequency vehicle target tracking algorism based on TLD innovatory algorithm |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109636834A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110675428A (en) * | 2019-09-06 | 2020-01-10 | 鹏城实验室 | Target tracking method and device for human-computer interaction and computer equipment |
CN111462180A (en) * | 2020-03-30 | 2020-07-28 | 西安电子科技大学 | Object tracking method based on AND-OR graph AOG |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102982559A (en) * | 2012-11-28 | 2013-03-20 | 大唐移动通信设备有限公司 | Vehicle tracking method and system |
CN104574439A (en) * | 2014-12-25 | 2015-04-29 | 南京邮电大学 | Kalman filtering and TLD (tracking-learning-detection) algorithm integrated target tracking method |
US20180046188A1 (en) * | 2015-08-19 | 2018-02-15 | Eyedea Inc. | Unmanned aerial vehicle having automatic tracking function and method of controlling the same |
CN107705321A (en) * | 2016-08-05 | 2018-02-16 | 南京理工大学 | Moving object detection and tracking method based on embedded system |
CN108447079A (en) * | 2018-03-12 | 2018-08-24 | 中国计量大学 | A kind of method for tracking target based on TLD algorithm frames |
-
2018
- 2018-11-22 CN CN201811395493.8A patent/CN109636834A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102982559A (en) * | 2012-11-28 | 2013-03-20 | 大唐移动通信设备有限公司 | Vehicle tracking method and system |
CN104574439A (en) * | 2014-12-25 | 2015-04-29 | 南京邮电大学 | Kalman filtering and TLD (tracking-learning-detection) algorithm integrated target tracking method |
US20180046188A1 (en) * | 2015-08-19 | 2018-02-15 | Eyedea Inc. | Unmanned aerial vehicle having automatic tracking function and method of controlling the same |
CN107705321A (en) * | 2016-08-05 | 2018-02-16 | 南京理工大学 | Moving object detection and tracking method based on embedded system |
CN108447079A (en) * | 2018-03-12 | 2018-08-24 | 中国计量大学 | A kind of method for tracking target based on TLD algorithm frames |
Non-Patent Citations (1)
Title |
---|
龚小彪: ""基于TLD框架的目标跟踪算法研究"", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110675428A (en) * | 2019-09-06 | 2020-01-10 | 鹏城实验室 | Target tracking method and device for human-computer interaction and computer equipment |
CN111462180A (en) * | 2020-03-30 | 2020-07-28 | 西安电子科技大学 | Object tracking method based on AND-OR graph AOG |
CN111462180B (en) * | 2020-03-30 | 2023-03-24 | 西安电子科技大学 | Object tracking method based on AND-OR graph AOG |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Ke et al. | Multi-dimensional traffic congestion detection based on fusion of visual features and convolutional neural network | |
CN111932583A (en) | Space-time information integrated intelligent tracking method based on complex background | |
CN103971386B (en) | A kind of foreground detection method under dynamic background scene | |
US7929730B2 (en) | Method and system for object detection and tracking | |
CN108960047B (en) | Face duplication removing method in video monitoring based on depth secondary tree | |
CN115995063A (en) | Work vehicle detection and tracking method and system | |
Craciun et al. | Joint detection and tracking of moving objects using spatio-temporal marked point processes | |
CN105976397A (en) | Target tracking method based on half nonnegative optimization integration learning | |
Gao et al. | Feature Particles Tracking for Moving Objects. | |
CN104778699A (en) | Adaptive object feature tracking method | |
Ali et al. | Vehicle detection and tracking in UAV imagery via YOLOv3 and Kalman filter | |
CN109636834A (en) | Video frequency vehicle target tracking algorism based on TLD innovatory algorithm | |
Feng | Mask RCNN-based single shot multibox detector for gesture recognition in physical education | |
Kumar | Visual object tracking using deep learning | |
Wang et al. | Foreground Object Detecting Algorithm based on Mixture of Gaussian and Kalman Filter in Video Surveillance. | |
Douillard et al. | A spatio-temporal probabilistic model for multi-sensor object recognition | |
Kim et al. | Unsupervised moving object segmentation and recognition using clustering and a neural network | |
Yang et al. | Knowledge Distillation for Feature Extraction in Underwater VSLAM | |
Maga et al. | An efficient approach for object detection and tracking | |
Bai et al. | Pedestrian Tracking and Trajectory Analysis for Security Monitoring | |
Soeleman et al. | Tracking Moving Objects based on Background Subtraction using Kalman Filter | |
Wang et al. | Detection Anomaly in Video Based on Deep Support Vector Data Description. | |
Suri | Detection of Moving Vehicles on Highway using Fuzzy Logic for Smart Surveillance System | |
Zhou et al. | Human tracking by employing the scene information in underground coal mines | |
Foo et al. | A Fast-Compressive Tracking Integrated with Differential Evolution to Optimize Object Tracking Performance |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190416 |
|
RJ01 | Rejection of invention patent application after publication |