CN105976401B - Method for tracking target and system based on piecemeal multi-instance learning algorithm - Google Patents
Method for tracking target and system based on piecemeal multi-instance learning algorithm Download PDFInfo
- Publication number
- CN105976401B CN105976401B CN201610339958.2A CN201610339958A CN105976401B CN 105976401 B CN105976401 B CN 105976401B CN 201610339958 A CN201610339958 A CN 201610339958A CN 105976401 B CN105976401 B CN 105976401B
- Authority
- CN
- China
- Prior art keywords
- target
- classifier
- threshold
- tracking
- score
- 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
- 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
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20021—Dividing image into blocks, subimages or windows
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of method for tracking target based on piecemeal multi-instance learning algorithm, comprising: target image is divided into several block diagram photos；The Weak Classifier pond of each block diagram photo is obtained by multi-instance learning algorithm, and is selected several strong Weak Classifiers of classification capacity from Weak Classifier pond and constituted strong classifier；In object tracking process, the integrated classifier score of target image is calculated in conjunction with the strong classifier of all block diagram photos, and target position is determined according to calculated integrated classifier score.The problems such as above-mentioned method for tracking target performance of target tracking based on piecemeal multi-instance learning algorithm is higher, tracking process is stablized, and is able to solve serious illumination, pose variation and blocks.Invention additionally discloses a kind of Target Tracking Systems based on piecemeal multi-instance learning algorithm.
Description
Technical field
The present invention relates to digital image processing techniques field, more particularly to a kind of based on piecemeal multi-instance learning algorithm
Method for tracking target and system.
Background technique
Target following technology is one of the project that field of machine vision is concerned.In recent years, many scholars both domestic and external
It is dedicated to the research of target following technology and achieves some noticeable achievements.However, target following still faces many choose
War, such as: the problems such as noise, illumination, pose change, move mutation and block.To solve the above problems, Babenko proposition is based on
The method for tracking target of multi-instance learning (Multiple Instance Learning, MIL) algorithm.This method indicates sample
For by the packet of the tape label of multiple composition examples (positive closure or negative packet).When at least one example is timing, the packet in a packet
It is marked as positive closure.Conversely, the coating marks the packet that is negative when all examples are negative in a packet.Multi-instance learning algorithm
Combining target and background information (example in positive closure and negative packet) training obtain identification and classification device, and utilize gained identification and classification device
Target is separated from background.
However, MIL algorithm calculates, time-consuming is big, and the drift of Yi Fasheng tracking result, performance of target tracking is not high enough, and can not solve
The problems such as certainly serious illumination, pose change and block.
Summary of the invention
The technical problem to be solved by the present invention is in view of the above shortcomings of the prior art, provide a kind of performance of target tracking more
High, the tracking more stable method for tracking target and system based on piecemeal multi-instance learning algorithm of process.
In order to solve the above technical problems, the technical solution used in the present invention is: a kind of calculated based on piecemeal multi-instance learning
The method for tracking target of method, comprising the following steps:
Target image is divided into several block diagram photos；
Obtain the Weak Classifier pond of each block diagram photo by multi-instance learning algorithm, and from the Weak Classifier
Several strong Weak Classifiers of classification capacity are selected in pond constitutes strong classifier；
In object tracking process, the compressive classification of target image is calculated in conjunction with the strong classifier of all block diagram photos
Device score, and target position is determined according to calculated integrated classifier score.
Preferably, using the principle for the inner product for maximizing Weak Classifier and maximum likelihood probability from the Weak Classifier pond
It selects several strong Weak Classifiers of classification capacity and constitutes strong classifier.
Preferably, the integrated classifier score of target image is calculated in conjunction with the strong classifier of all block diagram photos
Method are as follows:
The classifier score of each block diagram photo is calculated according to corresponding strong classifier；
The average value for calculating each classifier score obtains the integrated classifier score of target image.
Preferably, as the integrated classifier score maximum of the candidate samples of target and when being greater than second threshold, mesh
No light, pose change and eclipse phenomena during mark tracking；
The integrated classifier score of candidate samples as target is maximum and is less than the second threshold greater than first
When threshold value, if the classifier score of the part block diagram photo is greater than the second threshold, target is blocked by other objects；
If the classifier score of all block diagram photos is respectively less than the second threshold greater than the first threshold, there are illumination
Change with pose；Wherein, the first threshold is less than the second threshold；
When the integrated classifier score of the candidate samples as target is less than the first threshold, and all described points
When the classifier score of block image sheet is respectively less than the first threshold, target following failure；
When the integrated classifier score in continuous multiple frames as the candidate samples of target is respectively less than the second threshold,
And the classifier score of all block diagram photos is when being respectively less than the second threshold, target following failure.
Preferably, study update is carried out to classifier parameters according to current tracking mode after target following success；
Wherein, when no light, pose change and eclipse phenomena in object tracking process, study turnover rate takes 0.5；Work as mesh
When changing during mark tracking there are illumination, pose, study turnover rate takes 0.85；When target is blocked by other objects, study
Turnover rate takes 0.25.
A kind of Target Tracking System based on piecemeal multi-instance learning algorithm, including target image division module, classifier
Processing module and target position determining module；
The target image division module, for target image to be divided into several block diagram photos；
The classifier processing module, for obtaining weak point of each block diagram photo by multi-instance learning algorithm
Class device pond, and select several strong Weak Classifiers of classification capacity from the Weak Classifier pond and constitute strong classifier；
The target position determining module is used in object tracking process, in conjunction with the strong of all block diagram photos
The integrated classifier score of classifier calculated target image, and target position is determined according to calculated integrated classifier score.
Preferably, the classifier processing module is using the principle for maximizing Weak Classifier with the inner product of maximum likelihood probability
Several strong Weak Classifiers of classification capacity are selected from the Weak Classifier pond constitutes strong classifier.
Preferably, the target determination module calculates target image in conjunction with the strong classifier of all block diagram photos
The method of integrated classifier score are as follows:
The classifier score of each block diagram photo is calculated according to corresponding strong classifier；
The average value for calculating each classifier score obtains the integrated classifier score of target image.
Preferably, as the integrated classifier score maximum of the candidate samples of target and when being greater than second threshold, institute
It states target position determining module and determines that no light in object tracking process, pose changes and eclipse phenomena；
The integrated classifier score of candidate samples as target is maximum and is less than the second threshold greater than first
When threshold value, if the classifier score of the part block diagram photo is greater than the second threshold, the target position determines mould
Block determines that target is blocked by other objects；If it is big that the classifier score of all block diagram photos is respectively less than the second threshold
In the first threshold, then the target position determining module determines that there are illumination and pose to change；
When the integrated classifier score of the candidate samples as target is less than the first threshold, and all described points
When the classifier score of block image sheet is respectively less than the first threshold, the target position determining module determines that target following is lost
It loses；
When the integrated classifier score in continuous multiple frames as the candidate samples of target is respectively less than the second threshold,
And the classifier score of all block diagram photos, when being respectively less than the second threshold, the target position determining module determines
Target following failure.
Preferably, the Target Tracking System based on piecemeal multi-instance learning algorithm further includes study update module；Institute
Study update module is stated for the classifier after target following success according to current tracking mode to the classification processing module
Parameter carries out study update；
Wherein, when no light, pose change and eclipse phenomena in object tracking process, study turnover rate takes 0.5；Work as mesh
When changing during mark tracking there are illumination, pose, study turnover rate takes 0.85；When target is blocked by other objects, study
Turnover rate takes 0.25.
The beneficial effects of adopting the technical scheme are that target image is divided into multiple block diagram photos, needle
To each block diagram photo application multi-instance learning algorithm training identification and classification device, and all piecemeals are combined in object tracking process
The integrated classifier score of image sheet determines target position, so as to further increase performance of target tracking.Further, according to
The integrated classifier score of the classifier score of each block diagram photo and all block diagram photos can also detect and distinguish with
Illumination, pose variation and occlusion issue during track.And the study turnover rate that setting is different, classifier can be according to difference
Situation realizes adaptive updates, to realize lasting, stable target following.
Detailed description of the invention
Fig. 1 is that the present invention is based on the flow diagrams of the method for tracking target of piecemeal multi-instance learning algorithm；
Fig. 2 is that the present invention is based on the Target Tracking System structural schematic diagrams of piecemeal multi-instance learning algorithm.
Specific embodiment
The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
The basic thought of existing multi-instance learning algorithm is to realize multi-instance learning under Online Boosting frame
Algorithm.The algorithm forms in target and background region acquisition example be used as training sample with markd packet respectively, it may be assumed that { (X1,
y1),…,(Xi,yi),…,(Xm,ym)}.Wherein, Xi={ xi1,xi2,,xinIt is by example { xi1,xi2,,xinComposition packet, yi
For the label of the packet.Work as yiThe packet is positive closure when=1, works as yiThe packet is negative packet when=0.
Positive closure is by the composition examples that acquire in the field of present frame target position:Its
InFor target current location, r is the radius of circle of acquisition example.The negative packet of acquisition example composition at background:Wherein r, β (r < β) are inside and outside the half of the exemplary annulus of the negative packet of acquisition composition
Diameter.The example in positive and negative packet is described with Haar-like feature, it may be assumed that V=(v1,…,vN)T.To exemplary feature in positive and negative packet
Study obtains N number of Weak Classifier, constitutes Weak Classifier pond φ=(h1,…hk,…,hN)T.Weak Classifier uses Bayesian formula
It calculates.
Assuming that all equal Gaussian distributeds of featureI.e. And meet p (y=1)=p (y=0).Then to exemplary kth all in positive and negative packet
A feature is learnt to obtain k-th of Weak Classifier are as follows:
According to max log likelihood probability principle in Weak Classifier pond the strong K of selection sort ability (K < < N) a weak typing
Device constitutes strong classifier, for tracking target in subsequent video images.That is:
L is log-likelihood probability in formula:
p(yi|Xi) it is packet probability, multi-instance learning algorithm uses Noisy-OR (NOR) model:
p(yi|xij) it is example probability in packet:
p(yi|xij)=σ (H (xij)) (5)
WhereinFor sigmoid function.
Several strong Weak Classifiers of the classification capacity chosen constitute strong classifier:
In new frame image, strong classifier candidate samples concentrate using the maximum sample of classifier score as finally with
Track target:
Wherein candidate samples collection selects in the field of previous frame tracking result, it may be assumed that
After tracing into target, it is as follows to update rule to adapt to the variation such as illumination, pose for online real-time update Weak Classifier:
Wherein, 0 < λ < 1 is learning rate,vk(xi) respectively positive and negative
The mean value of exemplary characteristics in packet.It is respectively
Exemplary variance in positive and negative packet.
Existing multi-instance learning algorithm solves the ambiguity problem in object tracking process, however its there are still certain
The shortcomings that: need to calculate packet probability and example probability M times when the strong Weak Classifier of each selection sort ability, therefore the algorithm is real-time
Property is poor；Classifier is updated using fixed learning rate, easily causes update insufficient when illumination occurring or pose changes, works as generation
Yi Fasheng " cross and update " when blocking.
Referring to Fig. 1, in one embodiment, the present invention is based on the method for tracking target of piecemeal multi-instance learning algorithm can wrap
Include following steps:
Target image is divided into several block diagram photos by S100.
Wherein, to meet tracking performance requirement and real-time, target image can be uniformly divided into 9 block diagram photos,
That is O={ o1,…,oi,…,o9}.Wherein, O represents target image, oiRepresent i-th of block diagram photo, 1≤i≤9.Certainly, exist
In other embodiments, target image can also be divided into the block diagram photo of other numbers, it is without limitation.
S200, obtains the Weak Classifier pond of each block diagram photo by multi-instance learning algorithm, and from Weak Classifier pond
In select several strong Weak Classifiers of classification capacity and constitute strong classifier.
Wherein, in each block diagram photo oiSmall neighbourhood within the scope of acquisition example form positive closure?
The negative packet of acquisition example composition in its annulusThe Weak Classifier of image sheet is obtained to example training in packetI=1 ..., 9, k=1 ..., N.K=2 (K < < N) a Weak Classifier is selected to constitute strong classifier H in Weak Classifier pondi,
(i=1 ..., 9) it is used for subsequent video images.
Preferably, use the principle for the inner product for maximizing Weak Classifier and maximum likelihood probability from Weak Classifier in this step
The strong Weak Classifier of classification capacity is selected in pond as strong classifier, needs repeatedly choosing when can be avoided selection Weak Classifier every time
Select packet probability and example probability.
In object tracking process, the synthesis point of target image is calculated in conjunction with the strong classifier of all block diagram photos by S300
Class device score, and target position is determined according to calculated integrated classifier score.
Wherein, the side of the integrated classifier score of target image is calculated in conjunction with the strong classifier of all block diagram photos
Method are as follows:
Firstly, calculating the classifier score of each block diagram photo according to corresponding strong classifier.
Then, the average value for calculating each classifier score obtains the integrated classifier score of target image.
In object tracking process, collecting sample constitutes candidate samples collection X in the target position neighborhood of previous frame images。
To each candidate samples piecemeal, corresponding strong classifier H is utilizediCalculate the block diagram photo classifier score of candidate samplesi
=1 ..., 9, j=1 ..., Ns, wherein NsFor candidate samples sum.For a candidate samples, point of its all image sheet is counted
Class device score obtains integrated classifier scoreThe maximum candidate of integrated classifier score is concentrated in candidate samples
SampleFor target.
Illumination, pose variation and occlusion issue in object tracking process affect tracking performance.In the present embodiment, according to
It is detected according to the integrated classifier score of target and the classifier score of each block diagram photo and distinguishes normal tracking, illumination, position
Appearance changes, occlusion issue and the tracking tracking modes such as unsuccessfully.Set two comparison threshold values, respectively first threshold th1With second
Threshold value th2, and first threshold th1Less than second threshold th2.Then target following state is in the following several ways:
1) as the integrated classifier score maximum of the candidate samples of target and greater than second threshold th2When, target following
No light, pose change and eclipse phenomena in the process.
2) as the integrated classifier score maximum of the candidate samples of target and less than second threshold th2Greater than first threshold
th1When (), if the classifier score of the part block diagram photo of the sample is greater than second threshold th2, this
When target blocked by other objects.Wherein, the classifier score of the block diagram photo of shield portions is less than second threshold th2Even
Less than first threshold th1。
3) as the integrated classifier score maximum of the candidate samples of target and less than second threshold th2Greater than first threshold
th1When, if the classifier score of all block diagram photos of the sample is respectively less than second threshold th2Greater than first threshold th1, then
There are illumination, poses to change.
4) when the integrated classifier score of the candidate samples as target traced into is less than first threshold th1, and it is all
The classifier score of block diagram photo is respectively less than first threshold th1When, target following failure.
5) when the integrated classifier score in continuous multiple frames as the candidate samples of target is respectively less than second threshold th2, and
The classifier score of its all block diagram photo is respectively less than second threshold th2When, target following failure.
It further,, can be according to current tracking after target following success to realize lasting, stable target following
State is updated classifier parameters.Existing multi-instance learning algorithm is with fixed learning rate undated parameter.When there are light
According to or pose change when, too small learning rate can be such that classifier updates not in time to cause subsequent tracking failure.When target quilt
When other objects block, excessive learning rate leads to the interference that shelter is introduced in classifier renewal process.
In the present embodiment, classifier adaptive updates method is proposed.After tracking successfully, not according to the setting of current tracking mode
Same learning rate.
Wherein, when the problems such as no light, pose change and eclipse phenomena in object tracking process, study turnover rate is taken
0.5, classifier relies on the target traced into and is updated while considering target template at this time.When there are light in object tracking process
When changing according to, pose, target appearance variation is more obvious, and study turnover rate takes 0.85, and classifier renewal process more relies on
In current tracking result to adapt to outward appearance change.When target is blocked by other objects, study turnover rate takes 0.25, to avoid drawing
Enter the feature of non-targeted object, classifier is mainly updated according to previous frame target signature.For in object tracking process
Different conditions are respectively adopted different study turnover rate and update classifier, solve illumination during tracking, pose variation and
Occlusion issue guarantees long-time, tenacious tracking target.
Target image is divided into multiple block images by the above-mentioned method for tracking target based on piecemeal multi-instance learning algorithm
Piece for each block diagram photo application multi-instance learning algorithm training identification and classification device, and combines institute in object tracking process
There is the integrated classifier score of block diagram photo to determine target position, so as to further increase performance of target tracking.Into one
Step, can also detect simultaneously according to the integrated classifier score of the classifier score of each block diagram photo and all block diagram photos
Distinguish illumination, pose variation and the occlusion issue during tracking.And setting study turnover rate, classifier can be according to difference
Situation realizes adaptive updates, to realize lasting, stable target following.
Based on the same inventive concept, the present invention also proposes a kind of target following system based on piecemeal multi-instance learning algorithm
System.Due to the Target Tracking System based on piecemeal multi-instance learning algorithm and the aforementioned mesh based on piecemeal multi-instance learning algorithm
It is identical to mark tracking realization principle, therefore overlaps will not be repeated.
Referring to fig. 2, in one embodiment, the Target Tracking System based on piecemeal multi-instance learning algorithm may include target
Image division module 100, classifier processing module 200 and target position determining module 300.Wherein, target image division module
100, for target image to be divided into several block diagram photos.Classifier processing module 200, for being calculated by multi-instance learning
Method obtains the Weak Classifier pond of each block diagram photo, and several strong weak typings of classification capacity are selected from Weak Classifier pond
Device constitutes strong classifier.Target position determining module 300 is used in object tracking process, in conjunction with all block diagram photos
Strong classifier calculates the integrated classifier score of target image, and determines target position according to calculated integrated classifier score
It sets.
In one embodiment, classifier processing module 200 is using the inner product for maximizing Weak Classifier and maximum likelihood probability
Principle select several strong Weak Classifiers of classification capacity from Weak Classifier pond and constitute strong classifier, can be avoided each choosing
Repeatedly selection packet probability and example probability are needed when selecting Weak Classifier.
Specifically, target determination module 200 combines the synthesis of the strong classifier calculating target image of all block diagram photos
The method of classifier score can be with are as follows: firstly, calculating the classifier score of each block diagram photo according to corresponding strong classifier；
Then, the average value for calculating each classifier score obtains the integrated classifier score of target image.
In addition, the illumination, pose variation and occlusion issue in object tracking process affect tracking performance.The present embodiment
In, it is detected according to the integrated classifier score of target and the classifier score of each block diagram photo and distinguishes normal tracking, light
Change according to, pose, occlusion issue and the tracking tracking modes such as unsuccessfully.Set two comparison threshold values, respectively first threshold th1With
Second threshold th2, and first threshold th1Less than second threshold th2.Then target following state is in the following several ways:
1) as the integrated classifier score maximum of the candidate samples of target and greater than second threshold th2When, target position
Determining module 400 determines that no light in object tracking process, pose changes and eclipse phenomena.
2) as the integrated classifier score maximum of the candidate samples of target and less than second threshold th2Greater than first threshold
th1When (), if the classifier score of the part block diagram photo of the sample is greater than second threshold th2, then
Target position determining module 400 determines that target is blocked by other objects.Wherein, the classifier of the block diagram photo of shield portions point
Number is less than second threshold th2Even less than first threshold th1。
3) as the integrated classifier score maximum of the candidate samples of target and less than second threshold th2Greater than first threshold
th1When, if the classifier score of all block diagram photos of the sample is respectively less than second threshold th2Greater than first threshold th1, then
Target position determining module 400 determines that there are illumination, poses to change.
4) when the integrated classifier score of the candidate samples as target traced into is less than first threshold th1, and it is all
The classifier score of block diagram photo is respectively less than first threshold th1When, target position determining module 400 determines that target following is lost
It loses.
5) when the integrated classifier score in continuous multiple frames as the candidate samples of target is respectively less than second threshold th2, and
The classifier score of its all block diagram photo is respectively less than second threshold th2When, target position determining module 400 determine target with
Track failure.
Further, referring to fig. 2, the Target Tracking System based on piecemeal multi-instance learning algorithm can also include learning more
New module 400.Learn update module 400 to be used for after target following success according to current tracking mode to classification processing module
200 classifier parameters carry out study update.Wherein, when no light, pose change and eclipse phenomena in object tracking process,
Study turnover rate takes 0.5.When changing in object tracking process there are illumination, pose, study turnover rate takes 0.85.When target quilt
When other objects block, study turnover rate takes 0.25.
Target image is divided into multiple block images by the above-mentioned Target Tracking System based on piecemeal multi-instance learning algorithm
Piece for each block diagram photo application multi-instance learning algorithm training identification and classification device, and combines institute in object tracking process
There is the integrated classifier score of block diagram photo to determine target position, so as to further increase performance of target tracking.Into one
Step, can also detect simultaneously according to the integrated classifier score of the classifier score of each block diagram photo and all block diagram photos
Distinguish illumination, pose variation and the occlusion issue during tracking.And setting study update module, enable to classifier root
Adaptive updates are realized according to different situations, to realize lasting, stable target following.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (8)
1. a kind of method for tracking target based on piecemeal multi-instance learning algorithm, which comprises the following steps:
Target image is divided into several block diagram photos；
The Weak Classifier pond of each block diagram photo is obtained by multi-instance learning algorithm, and from the Weak Classifier pond
It selects several strong Weak Classifiers of classification capacity and constitutes strong classifier；
In object tracking process, the integrated classifier point of target image is calculated in conjunction with the strong classifier of all block diagram photos
Number, and target position is determined according to calculated integrated classifier score；Strong point of all block diagram photos of combination
The method that class device calculates the integrated classifier score of target image are as follows: each block diagram is calculated according to corresponding strong classifier
The classifier score of photo；The average value for calculating each classifier score obtains the integrated classifier score of target image；It is described
Determine that target position includes: the integrated classifier score and each piecemeal according to target according to calculated integrated classifier score
The classifier score of image sheet detects and distinguishes the tracking shape of normal tracking, the change of illumination pose, occlusion issue and tracking failure
State.
2. the method for tracking target according to claim 1 based on piecemeal multi-instance learning algorithm, which is characterized in that use
If it is strong that the principle for maximizing the inner product of Weak Classifier and maximum likelihood probability selects classification capacity from the Weak Classifier pond
Dry Weak Classifier constitutes strong classifier.
3. the method for tracking target according to claim 1 based on piecemeal multi-instance learning algorithm, which is characterized in that as
The integrated classifier scores of the candidate samples of target is maximum and when being greater than second threshold, no light in object tracking process,
Pose changes and eclipse phenomena；
The integrated classifier score of candidate samples as target is maximum and is less than the second threshold greater than first threshold
When, if the classifier score of the part block diagram photo is greater than the second threshold, target is blocked by other objects；If institute
There is the classifier score of the block image piece to be respectively less than the second threshold greater than the first threshold, then there is illumination and position
Appearance changes；Wherein, the first threshold is less than the second threshold；
When the integrated classifier score of the candidate samples as target is less than the first threshold, and all block diagrams
When the classifier score of photo is respectively less than the first threshold, target following failure；
When the integrated classifier score in continuous multiple frames as the candidate samples of target is respectively less than the second threshold, and institute
When thering is the classifier score of the block image piece to be respectively less than the second threshold, target following failure.
4. the method for tracking target according to claim 3 based on piecemeal multi-instance learning algorithm, which is characterized in that target
Study update is carried out to classifier parameters according to current tracking mode after tracking successfully；
Wherein, when no light, pose change and eclipse phenomena in object tracking process, study turnover rate takes 0.05；Work as target
When changing during tracking there are illumination, pose, study turnover rate takes 0.85；When target is blocked by other objects, study is more
New rate takes 0.25.
5. a kind of Target Tracking System based on piecemeal multi-instance learning algorithm, which is characterized in that divide mould including target image
Block, classifier processing module and target position determining module；
The target image division module, for target image to be divided into several block diagram photos；
The classifier processing module, for obtaining the Weak Classifier of each block diagram photo by multi-instance learning algorithm
Pond, and select several strong Weak Classifiers of classification capacity from the Weak Classifier pond and constitute strong classifier；
The target position determining module is used in object tracking process, in conjunction with the strong classification of all block diagram photos
Device calculates the integrated classifier score of target image, and determines target position according to calculated integrated classifier score；It is described
Target position determining module calculates the integrated classifier score of target image in conjunction with the strong classifier of all block diagram photos
Method are as follows: the classifier score of each block diagram photo is calculated according to corresponding strong classifier；Calculate each classifier
The average value of score obtains the integrated classifier score of target image；The target position determining module is according to calculated comprehensive
It closes classifier score and determines that target position includes: according to the integrated classifier score of target and the classifier of each block diagram photo
Score detects and distinguishes the tracking mode of normal tracking, the change of illumination pose, occlusion issue and tracking failure.
6. the Target Tracking System according to claim 5 based on piecemeal multi-instance learning algorithm, which is characterized in that described
Classifier processing module is using the principle for the inner product for maximizing Weak Classifier and maximum likelihood probability from the Weak Classifier pond
It selects several strong Weak Classifiers of classification capacity and constitutes strong classifier.
7. the Target Tracking System according to claim 5 based on piecemeal multi-instance learning algorithm, which is characterized in that as
The integrated classifier scores of the candidate samples of target is maximum and when being greater than second threshold, and the target position determining module is sentenced
Set the goal no light during tracking, pose change and eclipse phenomena；
The integrated classifier score of candidate samples as target is maximum and is less than the second threshold greater than first threshold
When, if the classifier score of the part block diagram photo is greater than the second threshold, the target position determining module is sentenced
It sets the goal and is blocked by other objects；If the classifier score of all block diagram photos is respectively less than the second threshold greater than institute
First threshold is stated, then the target position determining module determines that there are illumination and pose to change；
When the integrated classifier score of the candidate samples as target is less than the first threshold, and all block diagrams
When the classifier score of photo is respectively less than the first threshold, the target position determining module determines target following failure；
When the integrated classifier score in continuous multiple frames as the candidate samples of target is respectively less than the second threshold, and institute
When having the classifier score of the block image piece to be respectively less than the second threshold, the target position determining module determines target
Tracking failure.
8. the Target Tracking System according to claim 7 based on piecemeal multi-instance learning algorithm, which is characterized in that described
Target Tracking System based on piecemeal multi-instance learning algorithm further includes study update module；The study update module is used for
Study update is carried out according to classifier parameters of the current tracking mode to the classifier processing module after target following success；
Wherein, when no light, pose change and eclipse phenomena in object tracking process, study turnover rate takes 0.05；Work as target
When changing during tracking there are illumination, pose, study turnover rate takes 0.85；When target is blocked by other objects, study is more
New rate takes 0.25.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610339958.2A CN105976401B (en) | 2016-05-20 | 2016-05-20 | Method for tracking target and system based on piecemeal multi-instance learning algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610339958.2A CN105976401B (en) | 2016-05-20 | 2016-05-20 | Method for tracking target and system based on piecemeal multi-instance learning algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105976401A CN105976401A (en) | 2016-09-28 |
| CN105976401B true CN105976401B (en) | 2019-03-12 |
Family
ID=56956135
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610339958.2A Active CN105976401B (en) | 2016-05-20 | 2016-05-20 | Method for tracking target and system based on piecemeal multi-instance learning algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105976401B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP6542824B2 (en) * | 2017-03-13 | 2019-07-10 | ファナック株式会社 | Image processing apparatus and image processing method for calculating likelihood of image of object detected from input image |
| CN107240120B (en) * | 2017-04-18 | 2019-12-17 | 上海体育学院 | Method and device for tracking moving target in video |
| CN108765383B (en) * | 2018-03-22 | 2022-03-18 | 山西大学 | Video description method based on deep migration learning |
| CN109767457B (en) * | 2019-01-10 | 2021-01-26 | 厦门理工学院 | Online multi-example learning target tracking method, terminal device and storage medium |
| CN111368917B (en) * | 2020-03-04 | 2023-06-09 | 西安邮电大学 | Multi-example integrated learning method for criminal investigation image classification |
| CN111476825B (en) * | 2020-03-10 | 2022-08-26 | 重庆邮电大学 | Anti-occlusion target tracking method based on multi-example learning and kernel correlation filter |
| CN117522925B (en) * | 2024-01-05 | 2024-04-16 | 成都合能创越软件有限公司 | Method and system for judging object motion state in mobile camera under attention mechanism |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101236608A (en) * | 2008-01-25 | 2008-08-06 | 清华大学 | Human face detection method based on picture geometry |
| CN102831129A (en) * | 2011-06-16 | 2012-12-19 | 富士通株式会社 | Retrieval method and system based on multi-instance learning |
| CN103325125A (en) * | 2013-07-03 | 2013-09-25 | 北京工业大学 | Moving target tracking method based on improved multi-example learning algorithm |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7986827B2 (en) * | 2006-02-07 | 2011-07-26 | Siemens Medical Solutions Usa, Inc. | System and method for multiple instance learning for computer aided detection |
-
2016
- 2016-05-20 CN CN201610339958.2A patent/CN105976401B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101236608A (en) * | 2008-01-25 | 2008-08-06 | 清华大学 | Human face detection method based on picture geometry |
| CN102831129A (en) * | 2011-06-16 | 2012-12-19 | 富士通株式会社 | Retrieval method and system based on multi-instance learning |
| CN103325125A (en) * | 2013-07-03 | 2013-09-25 | 北京工业大学 | Moving target tracking method based on improved multi-example learning algorithm |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105976401A (en) | 2016-09-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105976401B (en) | Method for tracking target and system based on piecemeal multi-instance learning algorithm | |
| CN109977798B (en) | Mask pooling model training and pedestrian re-identification method for pedestrian re-identification | |
| Li et al. | Robust visual tracking based on convolutional features with illumination and occlusion handing | |
| Leistner et al. | On robustness of on-line boosting-a competitive study | |
| All et al. | FlowBoost—Appearance learning from sparsely annotated video | |
| CN106096538A (en) | Face identification method based on sequencing neural network model and device | |
| CN107247956A (en) | A kind of fast target detection method judged based on grid | |
| Liang et al. | Tvparser: An automatic tv video parsing method | |
| CN110929679A (en) | Non-supervision self-adaptive pedestrian re-identification method based on GAN | |
| CN106296742A (en) | A kind of online method for tracking target of combination Feature Points Matching | |
| CN105931276B (en) | A kind of long-time face tracking method based on patrol robot intelligence cloud platform | |
| CN104036237A (en) | Detection method of rotating human face based on online prediction | |
| CN106023155A (en) | Online object contour tracking method based on horizontal set | |
| Schinas et al. | CERTH@ MediaEval 2012 Social Event Detection Task. | |
| Xiao et al. | Traffic sign detection based on histograms of oriented gradients and boolean convolutional neural networks | |
| CN104036238B (en) | The method of the human eye positioning based on active light | |
| Burceanu et al. | Learning a robust society of tracking parts using co-occurrence constraints | |
| Liang et al. | Deep correlation filter tracking with shepherded instance-aware proposals | |
| CN104517300A (en) | Vision judgment tracking method based on statistical characteristic | |
| Sternig et al. | Transientboost: On-line boosting with transient data | |
| Al-Behadili et al. | Incremental learning and novelty detection of gestures in a multi-class system | |
| CN112560651B (en) | Target tracking method and device based on combination of depth network and target segmentation | |
| Chen et al. | Robust anomaly detection via fusion of appearance and motion features | |
| Choi et al. | Deep manifold embedding active shape model for pose invarient face tracking | |
| CN114842189B (en) | Adaptive Anchor generation method for target detection |
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 |