CN105809718B - A kind of method for tracing object of track entropy minimization - Google Patents
A kind of method for tracing object of track entropy minimization Download PDFInfo
- Publication number
- CN105809718B CN105809718B CN201610141089.2A CN201610141089A CN105809718B CN 105809718 B CN105809718 B CN 105809718B CN 201610141089 A CN201610141089 A CN 201610141089A CN 105809718 B CN105809718 B CN 105809718B
- Authority
- CN
- China
- Prior art keywords
- detector
- target
- track
- image
- frame
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
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 present invention provides a kind of method for tracing object of track entropy minimization, belong to computer graphical, image technique field.Specifically comprise the following steps：Object selection selects and determines the target object to be tracked from initial pictures.Image inputs, and under real-time disposition, extraction acquires by camera and be stored in the video image of memory block, as will be into the input picture of line trace；In processed offline, the video file acquired is decomposed into the image sequence of multiple frame compositions, sequentially in time, extracts frame image one by one as input picture.Detector creates and number of detectors is no more than 5.The each detector of pursuit path is generated simultaneously to be detected target, track entropy minimization target positioning, detector updates, according to determining target trajectory, detector training process in similar, positive and negative sample is extracted respectively in target area and background area to each frame corresponding to target trajectory, updates all detectors.For video frequency monitoring system.
Description
Technical field
The invention belongs to computer vision, figure, image technique fields.
Background technology
Visual object tracking is one of most important component part in computer vision application, these applications include intelligence prison
Control, human-computer interaction, automatic control system etc..Purpose to image tracing is the condition in the positions and dimensions of given initial object
Under, automatically determine its positions and dimensions in next each frame.Although the research of related object tracking has been carried out several
10 years, many important progress were also achieved in recent years, but due to the complexity of real world, such as background interference, it is apparent
And illumination variation, image low quality, frame-skip etc. so that design can reach and the comparable tracking of human levels still ten
Divide difficulty.One ideal tracking must consider real-time, stability and the persistence of tracking simultaneously.
Tracking can generally be divided into two class of method of formation and diagnostic method at present.Method of formation regards tracking problem in area as
The most like object of Search/Track target in domain, and target usually passes through the basal orientation in a sub-spaces (or template sequence)
Duration set is expressed.Different from method of formation, diagnostic method regards tracking as one will track what target was distinguished from background
Classification problem.Diagnostic method utilizes the information from target and background simultaneously, is the main method of current research object tracking.At this
In a little methods, Duffner etc. propose it is a kind of describing and promote the detector of Hough transformation based on pixel, and combine based on foreground and
The probabilistic segmentation method of background world model is realized quickly to image tracing.And in order to carry out for a long time to image tracing,
The method that many researchers use self study should by using the positive sample and negative sample more new model near target location
Method can be adaptively adjusted tracking system and deacclimatize that new target is apparent and background variation.However these methods update with
It is difficult to avoid that the error message constantly accumulated in track systematic procedure, therefore is easy to happen the position for drifting about and being detached from real target objects
It sets.It is difficult to the contradiction taken into account, Babenko to alleviate on-time model update is faced during tracking stability and flexibility
All uncertain positive samples and negative sample are all put into bag Deng a kind of method of online more case-based learnings, this method is used
Learn so obtain one for tracking identification and classification model.Mahadevan etc. propose a kind of differentiation that biology inspires with
Track method, this method include for the bottom-up center of study and the differentiation conspicuousness of peripheral region and space transforms model,
For the feature attention model of feature selecting, and for the top-down conspicuousness model of target detection.Kalal etc. proposes one
Kind the P-N learning methods of grader are obtained by positive sample and negative sample on-line study, this method by tracing task be decomposed into
Track, three parts of study and detection.Tracking section provides training examples for the update of detector, and detector then fails in tracking
When reinitialize tracking section, therefore tracking section and detector are supported and are promoted mutually.This mechanism also referred to as relies on
The tracking of detection, it has good tracking performance for prolonged tracing task.For the acute variation problem of scene,
Gall etc. proposes Hough forest based on random forest, and detects target by Hough transformation with this.The it is proposeds such as Zhang utilize
Different adaptation rates, which combines multiple graders and devises a kind of entropy computational methods, merges all tracking results.Ma etc. will be tracked
Task-decomposing is translation to target object and size estimation, and improves the accuracy of tracking and steady using correlation filtering method
It is qualitative.However these methods are still difficult to adapt to more complicated tracking environmental, this is analyzed and handled to pursuit path
Provide a kind of possible solution.Lu etc. and Supancic etc. evaluate pursuit path using Dynamic Programming, including
The time domain degree of correlation between the confidence level and continuous position of each position in track is calculated, is then tracked according to the evaluation result amendment
Track simultaneously updates apparent model, to achieve the purpose that inhibit error propagation.Lee etc. is based on various features tracking target and generates
Then multiple pursuit paths calculate the reliability of each track by analyzing these forward direction tracks and corresponding back trajectca-rles, and
It therefrom selects optimal track as target trajectory, detects and handled tracking mistake to a certain extent, improve the steady of tracking
It is qualitative.
Invention content
The object of the present invention is to provide a kind of method for tracing object of track entropy minimization, it can be efficiently solved to target
The tracking problem of object long-time real-time stabilization.
The purpose of the present invention is achieved through the following technical solutions：Specifically comprise the following steps：
(1) Object selection
The target object to be tracked is selected and determined from initial pictures.Object selection process passes through moving object detection side
Method automatically extracts, or is specified manually by man-machine interaction method.
(2) image inputs
Under real-time disposition, extraction acquires by camera and is stored in the video image of memory block, as will be into
The input picture of line trace；In processed offline, the video file acquired is decomposed into the image sequence of multiple frame compositions
Row extract frame image as input picture, if input picture is sky, tracking terminates one by one sequentially in time.
(3) detector creates
If number of image frames divided byRemainder be 1, then according to current frame image create a new detector.This
In using random fern as basic detector, binary feature of 2 comparative features as detector is used in combination.Feature set is 200
2 comparative features generated at random when initial, by randomly choosing M=25 structure detector therein.Each detector packet
Containing B=5 fern, each fern includes D=5 feature.Then in target area and background area respectively extraction and target sizes
Identical positive sample and negative sample, and detector is trained and is updated after their rules are turned to 25 × 25 pixel sizes,
Here sample is image block.If the total number of detectorMore than 5, then earliest created detector is deleted, i.e. detector
Number is no more than 5.
(4) pursuit path is generated
Each detector is simultaneously detected target, and using optimal testing result as where target in present frame
Then position generates respective pursuit path, the length of track in a manner of continuously detectingFrame.
(5) entropy minimization target in track positions
For the same target, different detectors will generate different pursuit paths, while correspondence obtains different mesh
Cursor position.IfIndicate detectors set,Indicate FjLoss function, wherein FjIndicate j-th of detector, N
For number of detectors, then optimal detector is calculated as：
That is detector of the selection with least disadvantage value.
Use PLL (Partial-Label Learning) method to solve loss function hereI.e. by maximizing mould
The posterior probability of shape parameter marks problem concerning study to solve part.IfIndicate a track set,For i-th of pursuit path,For XiIncluding period in kth frame position, H is the track frame number that includes
Mesh.If yi=(ci,vi) indicate XiCompound token, ci∈ { 1,0 } is that track marks (ci=1 indicates xiFor target trajectory, ci=0
Indicate XiFor non-targeted track), viFor XiTrailing end position.It is only possible in S there are one pursuit path be the true rail of target
Mark, the compound token collection of the target real trace of SIt is contained in possible compound token collectionIn,
Wherein to eachHaveAnd as i=k,For tracking problem,
F for detector selectionjLoss function can be calculated as：
φFj(S, Z)=- L (Fj；S,Z)+λH(Y|S,Z；Fj),
Wherein, L (Fj；S, Z) it indicates with FjFor the log-likelihood probability of model parameter, and H (Y | S, Z；Fj) indicate about training
The type mark empirical condition entropy of data and tag set, λ are the proportionality coefficient of the two, here λ=10.For tag set,
With relatively low probabilistic model parameter, its corresponding entropy is also smaller.That is, if one of model is marked
Note has higher probability and another marking probability is smaller, then the entropy of the model is also smaller；And if two of a model
Label probability all having the same, then the entropy of the model is larger.Here log-likelihood is defined as：
And entropy is calculated as：
p(Y|S；Fj) can be calculated as：
Wherein p (vi|ci)=p (vi|ci,Xi) be track spatial prior probabilities, its value be equal to track in each position
The average value of the detector class probability of correspondence image block, p (ci|xi；Fj) be track posterior probability.p(vi|ci) it is calculated as rail
The average value of the detector class probability of each position correspondence image block in mark, i.e.,WhereinFor inspection
Class probability of the survey device in the track kth frame.And p (ci|xi；Fj) then reliable by calculating the track based on image block Duplication
Degree obtains, and is calculated as：
WhereinFor forward trace trackWith its traceback trackIn kth frame
The image block Duplication at place, is calculated as：
WithIt indicates respectivelyWithImage block area at kth frame,It indicatesWith
The overlapping area of two image blocks at kth frame.Finally, p (Y | S, Z；Fj) be calculated as about p (Y | S；Fj) Kullback-
Leibler is projected：
Wherein, if Y ∈ Z, δZ(Y)=1, otherwise δZ(Y)=0, Ψ indicates Y, the vector space where Z.From there through
It calculates the loss function of each detector and select the detector with least disadvantage value as optimum detector, and by the detection
The corresponding pursuit path of device completes target positioning as target trajectory.
(6) detector updates
According to the target trajectory that (5) determine, according to the detector training process in (3), to each corresponding to target trajectory
Frame extracts positive sample respectively in target area and background area and negative sample updates all detectors.It jumps to (2).
For the present invention during tracking, the multiple detectors created online are used to record the apparent information of history of target, it
Respective pursuit path is generated by way of continuously detecting, be then based on track entropy minimization analysis fusion these tracking knots
Fruit simultaneously therefrom selects optimal detector, finally completes target positioning using its corresponding pursuit path as target trajectory, in turn
Realize tracking.
The present invention compared with prior art the advantages of and good effect：
This method creates multiple detectors online, and each detector generates respective tracking rail by the way of continuously detecting
Then mark selects optimal detector by the analysis of track entropy minimization, and then determines target trajectory, realize and appoint to image tracing
Business.Since this method creates multiple detectors to adapt to the variation of target during tracking in different moments, while passing through rail
The tracking result of all detectors is merged in the analysis of mark entropy minimization, and therefrom selects optimal detector to determine target trajectory,
And then the stability of tracking is improved, it can realize the target following of long-time real-time stabilization.
Description of the drawings
Fig. 1 is the techniqueflow chart of the present invention
Embodiment：
By taking the detection of highway video monitoring overspeed of vehicle as an example, tracking proposed by the present invention may be used and realize.
Specifically, it first by the background modeling and foreground extracting method that are widely used at present, obtains each in video monitoring range
The image-region of a vehicle, then using these image-regions as target object into line trace.To each such vehicle mesh
Mark, according to the method for the present invention, creates multiple detectors online first, is then continuously detected to target simultaneously, to generate
Respective pursuit path then will be with least disadvantage value then according to the penalty values of each detector of these trajectory calculations
Detector completes the positioning to vehicle, Jin Ershi as optimum detector, and using its corresponding pursuit path as target trajectory
Existing vehicle tracking.Finally, according to the image distance of the result calculating vehicle target movement in 1 second of vehicle target tracking, and according to
The actual motion distance of vehicle in the road is calculated in the proportionate relationship of image distance and actual range, and then obtains vehicle
Travel speed, if car speed has been more than the speed limit of highway, then it is assumed that the vehicle has exceeded the speed limit, and completes overspeed of vehicle inspection
It surveys.
The method of the present invention can be additionally used in the other application occasion to image tracing, such as intelligent video analysis, and human-computer interaction is handed over
Logical video monitoring, vehicle drive, and biocenose analysis and flow surface test the speed.
The method of the present invention can be programmed by any computer programming language (such as C language) and be realized, based on this method
Tracking system software can realize real-time objects tracking application in any PC or embedded system.
Claims (1)
1. a kind of method for tracing object of track entropy minimization, includes the following steps：
(1) Object selection
Select and determine the target object to be tracked from initial pictures, Object selection process by moving target detecting method from
Dynamic extraction, or specified manually by man-machine interaction method；
(2) image inputs
Under real-time disposition, extraction acquires by camera and is stored in the video image of memory block, as to carry out with
The input picture of track；In processed offline, the video file acquired is decomposed into the image sequence of multiple frame compositions, is pressed
According to time sequencing, frame image is extracted one by one as input picture, if input picture is sky, tracking terminates；
(3) detector creates
If number of image frames divided byRemainder be 1, then according to current frame image create a new detector, adopt here
It uses random fern as basic detector, binary feature of 2 comparative features as detector is used in combination；Feature set is 200 initial
2 comparative features that Shi Suiji is generated, by randomly choosing M=25 structure detector therein, each detector includes B
=5 ferns, each fern include D=5 feature；Then it is extracted respectively in target area and background area identical as target sizes
Positive sample and negative sample, and detector is trained and is updated after their rules are turned to 25 × 25 pixel sizes, here
Sample is image block, if the total number of detectorMore than 5, then earliest created detector is deleted, i.e. number of detectors
No more than 5；
(4) pursuit path is generated
Each detector is simultaneously detected target, and using optimal testing result as the position where target in present frame
It sets, respective pursuit path, the length of track is then generated in a manner of continuously detectingFrame；
(5) entropy minimization target in track positions
For the same target, different detectors will generate different pursuit paths, while correspondence obtains different target positions
It sets, ifIndicate detectors set,Indicate FjLoss function, wherein FjIndicate that j-th of detector, N are inspection
Device number is surveyed, then optimal detector is calculated as：
That is detector of the selection with least disadvantage value；
Use Partial-Label Learning methods to solve loss function hereI.e. by maximize model parameter after
It tests probability and marks problem concerning study to solve part, ifIndicate a track set,It is i-th
A pursuit path,For XiIncluding period in kth frame position, H is the track frame number that includes, if yi=(ci,vi) indicate
XiCompound token, ci∈ { 1,0 } marks for track, ci=1 indicates xiFor target trajectory, ci=0 indicates XiFor non-targeted track,
viFor XiTrailing end position, be only possible in S there are one pursuit path be target real trace, the target real trace of S is answered
Close label setsIt is contained in possible compound token collectionIn, wherein to eachHaveAnd as i=k,For tracking problem, selected for detector
The F selectedjLoss function can be calculated as：
Wherein, L (Fj；S, Z) it indicates with FjFor the log-likelihood probability of model parameter, and H (Y | S, Z；Fj) indicate about training data
With the type mark empirical condition entropy of tag set, λ is the proportionality coefficient of the two, and λ=10 have tag set here
Its corresponding entropy of relatively low probabilistic model parameter is also smaller, that is to say, that if one of model label tool
There is higher probability and another marking probability is smaller, then the entropy of the model is also smaller；And if two labels of a model
Probability all having the same, then the entropy of the model is larger, and log-likelihood is defined as here：
And entropy is calculated as：
p(Y|S；Fj) can be calculated as：
Wherein p (vi|ci)=p (vi|ci,Xi) be track spatial prior probabilities, its value is equal to each position in track and corresponds to
The average value of the detector class probability of image block, p (ci|xi；Fj) be track posterior probability；p(vi|ci) be calculated as in track
The average value of the detector class probability of each position correspondence image block, i.e.,WhereinFor detector
In the class probability of the track kth frame, and p (ci|xi；Fj) then obtained by calculating the track reliability based on image block Duplication
It arrives, is calculated as：
WhereinFor forward trace trackWith its traceback trackAt kth frame
Image block Duplication, is calculated as：
WithIt indicates respectivelyWithImage block area at kth frame,It indicatesWith
The overlapping area of two image blocks at k frames；Finally, p (Y | S, Z；Fj) be calculated as about p (Y | S；Fj) Kullback-
Leibler is projected：
Wherein, if Y ∈ Z, δZ(Y)=1, otherwise δZ(Y)=0, Ψ indicates Y, the vector space where Z, from there through calculating
The loss function of each detector and select the detector with least disadvantage value as optimum detector, and by the detector pair
The pursuit path answered completes target positioning as target trajectory；
(6) detector updates
The target trajectory determined according to (5) exists to each frame corresponding to target trajectory according to the detector training process in (3)
Positive sample is extracted in target area and background area respectively and negative sample updates all detectors, is jumped to (2).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610141089.2A CN105809718B (en) | 2016-03-14 | 2016-03-14 | A kind of method for tracing object of track entropy minimization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610141089.2A CN105809718B (en) | 2016-03-14 | 2016-03-14 | A kind of method for tracing object of track entropy minimization |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105809718A CN105809718A (en) | 2016-07-27 |
| CN105809718B true CN105809718B (en) | 2018-07-27 |
Family
ID=56468295
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610141089.2A Expired - Fee Related CN105809718B (en) | 2016-03-14 | 2016-03-14 | A kind of method for tracing object of track entropy minimization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105809718B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106530324A (en) * | 2016-10-21 | 2017-03-22 | 华中师范大学 | Visual cortex mechanism simulated video object tracking method |
| CN106842625B (en) * | 2017-03-03 | 2020-03-17 | 西南交通大学 | Target tracking method based on feature consensus |
| CN108332751B (en) * | 2018-01-08 | 2020-11-20 | 北京邮电大学 | Multi-source fusion positioning method and device, electronic equipment and storage medium |
| CN109309864B (en) * | 2018-08-08 | 2019-06-07 | 艾博特(上海)电信科技有限公司 | Nationality's information intelligent identifying system |
| CN109272036B (en) * | 2018-09-14 | 2021-07-02 | 西南交通大学 | Random fern target tracking method based on depth residual error network |
| CN111965614B (en) * | 2020-07-27 | 2022-08-19 | 北京理工大学 | Maneuvering weak target detection method based on dynamic programming and minimum image entropy |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103208190A (en) * | 2013-03-29 | 2013-07-17 | 西南交通大学 | Traffic flow detection method based on object detection |
| CN104898116A (en) * | 2015-06-25 | 2015-09-09 | 苏州途视电子科技有限公司 | Phase-coherent accumulation tracking-before-detection method based on minimum entropy |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8970589B2 (en) * | 2011-02-10 | 2015-03-03 | Edge 3 Technologies, Inc. | Near-touch interaction with a stereo camera grid structured tessellations |
-
2016
- 2016-03-14 CN CN201610141089.2A patent/CN105809718B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103208190A (en) * | 2013-03-29 | 2013-07-17 | 西南交通大学 | Traffic flow detection method based on object detection |
| CN104898116A (en) * | 2015-06-25 | 2015-09-09 | 苏州途视电子科技有限公司 | Phase-coherent accumulation tracking-before-detection method based on minimum entropy |
Non-Patent Citations (5)
| Title |
|---|
| Robust object tracking using enhanced random ferns;Wei Quan 等;《Visual Computer》;20141231;全文 * |
| 可视对象跟踪算法研究及应用;权伟;《中国博士学位论文全文数据库 信息科技辑》;20141015;全文 * |
| 在线学习多重检测的可视对象跟踪方法;权伟;《电子学报》;20140531;第42卷(第5期);全文 * |
| 基于信息熵的Mean-shift目标跟踪;周尚波 等;《计算机系统应用》;20101231;第19卷(第12期);全文 * |
| 基于霍夫蕨的实时对象跟踪方法;权伟 等;《西南交通大学学报》;20140630;第49卷(第3期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105809718A (en) | 2016-07-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105809718B (en) | A kind of method for tracing object of track entropy minimization | |
| CN106169188B (en) | A kind of method for tracing object based on the search of Monte Carlo tree | |
| Xu et al. | Segment as points for efficient online multi-object tracking and segmentation | |
| Choi et al. | A general framework for tracking multiple people from a moving camera | |
| Lookingbill et al. | Reverse optical flow for self-supervised adaptive autonomous robot navigation | |
| Yang et al. | Multi-object tracking with discriminant correlation filter based deep learning tracker | |
| Wojek et al. | Monocular visual scene understanding: Understanding multi-object traffic scenes | |
| CN108830171B (en) | Intelligent logistics warehouse guide line visual detection method based on deep learning | |
| US9805256B2 (en) | Method for setting a tridimensional shape detection classifier and method for tridimensional shape detection using said shape detection classifier | |
| JP2021530048A (en) | Multi-layered target classification method and device, traffic sign detection method and device, device and medium | |
| CN107146237B (en) | Target tracking method based on online state learning and estimation | |
| CN102722725A (en) | Object tracing method based on active scene learning | |
| CN102799900A (en) | Target tracking method based on supporting online clustering in detection | |
| CN104615986A (en) | Method for utilizing multiple detectors to conduct pedestrian detection on video images of scene change | |
| CN108038515A (en) | Unsupervised multi-target detection tracking and its storage device and camera device | |
| Munir et al. | LDNet: End-to-end lane marking detection approach using a dynamic vision sensor | |
| CN109993770A (en) | A kind of method for tracking target of adaptive space-time study and state recognition | |
| CN113361533A (en) | Segmentation identification method and system for overlapped and shielded target object | |
| CN104778699A (en) | Adaptive object feature tracking method | |
| CN104637052A (en) | Object tracking method based on target guide significance detection | |
| Rosales et al. | Faster r-cnn based fish detector for smart aquaculture system | |
| Xiang et al. | Multitarget tracking using hough forest random field | |
| CN109493370A (en) | A kind of method for tracking target based on spatial offset study | |
| CN116434150B (en) | Multi-target detection tracking method, system and storage medium for congestion scene | |
| CN113129336A (en) | End-to-end multi-vehicle tracking method, system and computer readable medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180727 Termination date: 20210314 |