CN103942536B - Multi-target tracking method of iteration updating track model - Google Patents
Multi-target tracking method of iteration updating track model Download PDFInfo
- Publication number
- CN103942536B CN103942536B CN201410136574.1A CN201410136574A CN103942536B CN 103942536 B CN103942536 B CN 103942536B CN 201410136574 A CN201410136574 A CN 201410136574A CN 103942536 B CN103942536 B CN 103942536B
- Authority
- CN
- China
- Prior art keywords
- target
- response
- model
- target detection
- track
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a multi-target tracking method of an iteration updating track model. The method comprises the steps that firstly, a target detector trained in advance carries out target detection on each image in a video sequence, the track model is initialized through the detection result of the first frame of image, the probability function of a Markov random field is updated through the track model, tracking track fragments are solved through a belief propagation algorithm, the track model is updated again by selecting the confidence tracking track fragment, and the final target track is obtained through iteration. According to the method, through the iteration updating of the track model, the challenge problems of target shielding, missing detection and false detection of the detector, similar targets and the like in the multi-target tracking technology can be solved well, and multi-target reliable and accurate tracking can be achieved.
Description
Technical field
The invention belongs to image procossing and computer vision field, and in particular to a kind of iteration updates many mesh of locus model
Mark tracking.
Background technology
Multiple target tracking refer to video sequence in multiple interesting targets, maintain the identity of each target, and reasoning
The movable informations such as the position of target, speed.Multiple target tracking is an important subject of computer vision field, in intelligence
The numerous areas such as video monitoring, robotics, behavior analysiss have important using value.
Compare with monotrack task, multiple target tracking has more challenging difficult points.First, the number of target is not
Know, and as the number that target passes in and out scene objects can change.Secondly, target Jing is often by foreground object or other targets
Partial occlusion is even blocked completely, is caused algorithm keeps track failure or is caused the identity between target to exchange.Finally, when in scene simultaneously
There is the target of similar appearance, especially they are close to each other when blocking, it is difficult to distinguish each target.
With the progress of target detection technique, nearest multi-object tracking method is mainly with the track side based on detection
Method, i.e., using the object detector that a training in advance is good, carry out target detection, track algorithm in video sequence each image
Data cube computation is carried out to the detection response in time serieses mainly, the detection for belonging to same target is responded and is connected into one
Track.Chinese patent such as in Publication No. CN101339608A disclose " it is a kind of based on detection method for tracking target and be
System ", the patent set up tracking object queue, and carry out present frame detection response and object queue according to position and yardstick
Match somebody with somebody, realize that target following and state update, enhance the real-time of tracking.But occur to overlap, separate or cosmetic variation in target
When, reliability is low, tracks unstable.
The content of the invention
It is an object of the invention to overcome the shortcoming of above-mentioned prior art, there is provided a kind of iteration updates many of locus model
Method for tracking target, the tracking in high precision, reliably can be tracked to multiple targets.
To reach above-mentioned purpose, iteration of the present invention updates the multi-object tracking method of locus model includes following step
Suddenly:
1) each image performance objective in video sequence is detected by training in advance good object detector, obtains target
Detection response, obtains the Feature Descriptor of each target detection response, video sequence is equally divided into N number of time window then;
2) setting target trajectory model includes the Feature Descriptor at response target each moment in current time window, according to video
The target detection of the first frame of sequence responds to obtain target trajectory model T={ T1..., TK, wherein, TkThe rail of target is responded for k-th
Mark model, K are the quantity of response target in the first frame, and each locus model for responding target corresponds to a target detection and rings
Should, each position of the response target in whole time window is gone out according to target trajectory model assessment then;
3) observation during each target detection in each time window is responded as markov random file, each observation are set
One labelling of connection, sets up markov random file, then by the probability of target trajectory model modification markov random file
Function and smooth probability function, and the marginal probability distribution that each target detection response belongs to each locus model is obtained, work as target
When detection response belongs to the marginal probability of the locus model more than predetermined threshold value, then the labelling of target detection response is revised as
The target sequence number of respective response target;
4) target sequence number identical and from consecutive frame target detection response is coupled together, forms a target trajectory
Fragment, then increases to the response target for newly obtaining in target trajectory model, and deletes the corresponding rail of response target of disappearance
Mark model;
5) then repeat step 3) and step 4), obtain it is all response target trajectory pieces of the target in whole video sequence
Then the target trajectory fragment of same target sequence number is coupled together by section, then each response mesh is obtained after smoothed and interpolation processing
Target track.
Step 1) in the response of each target detection Feature Descriptor include responding target's center's point coordinates, speed, color it is straight
Side's figure and size.
Step 3) concretely comprise the following steps:If the target detection response in each time window is one of markov random file
Observation yi, each observation one labelling l of connectioni, then maximize markov random file conditional probability be:
Wherein, Z is normalization factor, and observation set of the Y for random field, L are the tag set of random field, and T is track mould
Type, φ (li, yi) for single-point probability function, ψ (li, lj) for smooth probability function;
By the single-point probability function φ (l of target trajectory model modification markov random filei, yi) and smooth probability letter
Number ψ (li, lj), recycle faith propagation algorithm to obtain the marginal probability distribution that each target detection response belongs to each response target,
When the marginal probability that target detection response belongs to the response target is more than predetermined threshold value a, then target detection is responded corresponding
Labelling is revised as the target sequence number of respective response target.
The invention has the advantages that:
Iteration of the present invention updates the multi-object tracking method of locus model when the track of multiple targets is obtained, first
The each image performance objective in video sequence is detected by training in advance good object detector, then by video sequence point
For N number of time window, and target trajectory model is obtained according to the target detection response of the first frame in each time window, then again by repeatedly
Substitute performance locus model updates and maximizes markov random file conditional probability, makes the more accurate of target trajectory model change, from
And accurate target trajectory is obtained, while keeping dbjective state to remain unchanged at short notice, so as to improve target following
Robustness.
Description of the drawings
Fig. 1 is the schematic flow sheet of the present invention.
Specific embodiment
Below in conjunction with the accompanying drawings the present invention is described in further detail:
With reference to Fig. 1, iteration of the present invention updates the multi-object tracking method of locus model, comprises the following steps:
1) each image performance objective in video sequence is detected by training in advance good object detector, obtains target
Detection response, obtains the Feature Descriptor of each target detection response, video sequence is equally divided into N number of time window then;
2) setting target trajectory model includes the Feature Descriptor at response target each moment in current time window, according to video
The target detection of the first frame of sequence responds to obtain target trajectory model T={ T1..., TK, wherein, TkTarget is responded for k-th
Locus model, K are the quantity of response target in the first frame, and each locus model for responding target corresponds to a target detection and rings
Should, each position of the response target in whole time window is gone out according to target trajectory model assessment then;
Wherein, locus model TkComprising following parameter: Represent target trajectory model
Current time window is predicted in each time of day response target location,The average speed of response target is represented,Represent
The color histogram of response target,Represent the size of response target.
3) observation during each target detection in each time window is responded as markov random file, each observation are set
One labelling of connection, sets up markov random file, then by the probability of target trajectory model modification markov random file
Function and smooth probability function, and the marginal probability distribution that each target detection response belongs to each locus model is obtained, work as target
When detection response belongs to the marginal probability of the locus model more than predetermined threshold value, then the labelling of target detection response is revised as
The target sequence number of respective response target;
4) target sequence number identical and from consecutive frame target detection response is coupled together, forms a target trajectory
Fragment, then increases to the response target for newly obtaining in target trajectory model, and deletes the corresponding rail of response target of disappearance
Mark model;
5) then repeat step 3) and step 4), obtain it is all response target trajectory pieces of the target in whole video sequence
Then the target trajectory fragment of same target sequence number is coupled together by section, then each response mesh is obtained after smoothed and interpolation processing
Target track.
Step 1) in the response of each target detection Feature Descriptor include responding target's center's point coordinates, speed, color it is straight
Side's figure and size.
Step 3) concretely comprise the following steps:If the target detection response in each time window is one of markov random file
Observation yi, each observation one labelling l of connectioni, then maximize markov random file conditional probability be:
Wherein, Z is normalization factor, and observation set of the Y for random field, L are the tag set of random field, and T is track mould
Type, φ (li, yi) for single-point probability function, ψ (li, lj) for smooth probability function;
By the single-point probability function φ (l of target trajectory model modification markov random filei, yi) and smooth probability letter
Number ψ (li, lj), recycle faith propagation algorithm to obtain the marginal probability distribution that each target detection response belongs to each response target,
When the marginal probability that target detection response belongs to the response target is more than predetermined threshold value a, then target detection is responded corresponding
Labelling is revised as the target sequence number of respective response target.
Wherein, the single-point probability function φ (l of k-th targeti, yi), it is defined as detecting response with k-th object module
Similarity, Similarity Measure are divided into center point coordinate, speed, color histogram and size.Smooth probability function ψ (li, lj) ask
Solution is directed to the adjacent node in neighborhood, including in front and back's frame adjacent node and the adjacent node for belonging to same frame.It is adjacent in front and back's frame
Node, when detection response, to belong to same target-like probability of state larger, otherwise less;Adjacent node to same frame, because mesh
Mark simultaneously can not possibly occur in piece image two it is local, so detection response belong to same target-like probability of state compared with
It is little, it is then larger on the contrary.
Claims (2)
1. a kind of iteration updates the multi-object tracking method of locus model, it is characterised in that comprise the following steps:
1) each image performance objective in video sequence is detected by training in advance good object detector, obtains target detection
Response, obtains the Feature Descriptor of each target detection response, video sequence is equally divided into N number of time window then;
2) setting target trajectory model includes the Feature Descriptor at response target each moment in current time window, according to video sequence
The target detection of the first frame responds to obtain target trajectory model T={ T1..., TK, wherein, TkThe track of target is responded for k-th
Model, K are the quantity of response target in the first frame, and each locus model for responding target corresponds to target detection response,
Then each position of the response target in whole time window is gone out according to target trajectory model assessment;
3) observation during each target detection in each time window is responded as markov random file is set, each observation connection
One labelling, sets up markov random file, then by the probability function of target trajectory model modification markov random file
And smooth probability function, and the marginal probability distribution that each target detection response belongs to each locus model is obtained, work as target detection
When response belongs to the marginal probability of the locus model more than predetermined threshold value, then the labelling of target detection response is revised as accordingly
The target sequence number of response target;
4) target sequence number identical and from consecutive frame target detection response is coupled together, forms a target trajectory piece
Section, then increases to the response target for newly obtaining in target trajectory model, and deletes the corresponding track of response target of disappearance
Model;
5) then repeat step 3) and step 4), obtain it is all response target trajectory fragments of the target in whole video sequence, so
Afterwards the target trajectory fragment of same target sequence number is coupled together, then after smoothed and interpolation processing, obtains the rail of each response target
Mark;
Step 1) in the Feature Descriptor of each target detection response include responding target's center's point coordinates, speed, color histogram
And size.
2. iteration according to claim 1 updates the multi-object tracking method of locus model, it is characterised in that step 3)
Concretely comprise the following steps:If the target detection response in each time window is an observation y of markov random filei, each observation
One labelling l of connectioni, then maximize markov random file conditional probability be:
Wherein, Z is normalization factor, observation set of the Y for random field, tag sets of the L for random field, and T is locus model, φ
(li, yi) for single-point probability function, ψ (li, lj) for smooth probability function;
By the single-point probability function φ (l of target trajectory model modification markov random filei, yi) and smooth probability function ψ
(li, lj), recycle faith propagation algorithm to obtain the marginal probability distribution that each target detection response belongs to each response target, when
When target detection response belongs to the marginal probability of the response target more than predetermined threshold value a, then target detection is responded into corresponding mark
Note is revised as the target sequence number of respective response target.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410136574.1A CN103942536B (en) | 2014-04-04 | 2014-04-04 | Multi-target tracking method of iteration updating track model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410136574.1A CN103942536B (en) | 2014-04-04 | 2014-04-04 | Multi-target tracking method of iteration updating track model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103942536A CN103942536A (en) | 2014-07-23 |
CN103942536B true CN103942536B (en) | 2017-04-26 |
Family
ID=51190200
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410136574.1A Active CN103942536B (en) | 2014-04-04 | 2014-04-04 | Multi-target tracking method of iteration updating track model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103942536B (en) |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105678804A (en) * | 2016-01-06 | 2016-06-15 | 北京理工大学 | Real-time on-line multi-target tracking method by coupling target detection and data association |
CN105957105B (en) * | 2016-04-22 | 2018-10-02 | 清华大学 | The multi-object tracking method and system of Behavior-based control study |
CN106022220B (en) * | 2016-05-09 | 2020-02-28 | 北京河马能量体育科技有限公司 | Method for tracking multiple faces of participating athletes in sports video |
CN109478333A (en) * | 2016-09-30 | 2019-03-15 | 富士通株式会社 | Object detection method, device and image processing equipment |
AU2017418043B2 (en) | 2017-07-13 | 2020-05-21 | Beijing Voyager Technology Co., Ltd. | Systems and methods for trajectory determination |
CN107883963B (en) * | 2017-11-08 | 2020-02-14 | 大连大学 | Position prediction algorithm based on combination of IRWQS and fuzzy features |
CN108063802B (en) * | 2017-12-01 | 2020-07-28 | 南京邮电大学 | User position dynamic modeling optimization method based on edge calculation |
CN108804539B (en) * | 2018-05-08 | 2022-03-18 | 山西大学 | Track anomaly detection method under time and space double view angles |
CN109389134B (en) * | 2018-09-28 | 2022-10-28 | 山东衡昊信息技术有限公司 | Image processing method of monitoring information system of meat product processing production line |
CN111489377B (en) * | 2020-03-27 | 2023-11-10 | 北京迈格威科技有限公司 | Target tracking self-evaluation method and device |
CN112990154B (en) * | 2021-05-11 | 2021-07-30 | 腾讯科技(深圳)有限公司 | Data processing method, computer equipment and readable storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101339608A (en) * | 2008-08-15 | 2009-01-07 | 北京中星微电子有限公司 | Object tracking method and system based on detection |
CN101944234A (en) * | 2010-07-23 | 2011-01-12 | 中国科学院研究生院 | Multi-object tracking method and device driven by characteristic trace |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100013935A1 (en) * | 2006-06-14 | 2010-01-21 | Honeywell International Inc. | Multiple target tracking system incorporating merge, split and reacquisition hypotheses |
-
2014
- 2014-04-04 CN CN201410136574.1A patent/CN103942536B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101339608A (en) * | 2008-08-15 | 2009-01-07 | 北京中星微电子有限公司 | Object tracking method and system based on detection |
CN101944234A (en) * | 2010-07-23 | 2011-01-12 | 中国科学院研究生院 | Multi-object tracking method and device driven by characteristic trace |
Also Published As
Publication number | Publication date |
---|---|
CN103942536A (en) | 2014-07-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103942536B (en) | Multi-target tracking method of iteration updating track model | |
CN108470332B (en) | Multi-target tracking method and device | |
CN107145862B (en) | Multi-feature matching multi-target tracking method based on Hough forest | |
Zhang et al. | Online multi-person tracking by tracker hierarchy | |
Bae et al. | Robust online multiobject tracking with data association and track management | |
CN106778712B (en) | Multi-target detection and tracking method | |
CN105023278A (en) | Movable target tracking method and system based on optical flow approach | |
CN109598684B (en) | Correlation filtering tracking method combined with twin network | |
CN107133970B (en) | Online multi-target tracking method and device based on motion information | |
CN110288627B (en) | Online multi-target tracking method based on deep learning and data association | |
CN104008371A (en) | Regional suspicious target tracking and recognizing method based on multiple cameras | |
Zhang et al. | Robust object tracking via active feature selection | |
CN103729861A (en) | Multiple object tracking method | |
Zheng et al. | Adaptive edge-based mean shift for drastic change gray target tracking | |
Jang et al. | Object tracking based on an online learning network with total error rate minimization | |
CN106599918A (en) | Vehicle tracking method and system | |
CN104463909A (en) | Visual target tracking method based on credibility combination map model | |
Ryoo et al. | Observe-and-explain: A new approach for multiple hypotheses tracking of humans and objects | |
Gad et al. | Real-time lane instance segmentation using segnet and image processing | |
Qing et al. | A novel particle filter implementation for a multiple-vehicle detection and tracking system using tail light segmentation | |
CN103996207A (en) | Object tracking method | |
CN117630860A (en) | Gesture recognition method of millimeter wave radar | |
Chen et al. | Intrusion detection of specific area based on video | |
Walsh et al. | Detecting tracking failures from correlation response maps | |
CN108241837B (en) | Method and device for detecting remnants |
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 | ||
TR01 | Transfer of patent right |
Effective date of registration: 20210106 Address after: 221600 North Side of Xiaohe Road, Peixian Development Zone, Xuzhou City, Jiangsu Province Patentee after: Xuzhou Guolong Electric Power Parts Foundry Co.,Ltd. Address before: 710049 No. 28 West Xianning Road, Shaanxi, Xi'an Patentee before: XI'AN JIAOTONG University |
|
TR01 | Transfer of patent right |