CN108062764A - A kind of object tracking methods of view-based access control model - Google Patents
A kind of object tracking methods of view-based access control model Download PDFInfo
- Publication number
- CN108062764A CN108062764A CN201711243635.4A CN201711243635A CN108062764A CN 108062764 A CN108062764 A CN 108062764A CN 201711243635 A CN201711243635 A CN 201711243635A CN 108062764 A CN108062764 A CN 108062764A
- Authority
- CN
- China
- Prior art keywords
- tracking
- object tracking
- target
- access control
- detection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/262—Analysis of motion using transform domain methods, e.g. Fourier domain methods
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
- G06T7/251—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving models
-
- 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
- G06T2207/20048—Transform domain processing
- G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]
-
- 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)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention is a kind of long-time object tracking method of view-based access control model, and step includes:Input is used as each two field picture of mobile phone acquisition, target to be tracked is selected by subscriber frame, runs multiple dimensioned KCF trackers.Tracking creditability is calculated by peak sidelobe ratio during tracking, tracking filter model is updated in high confidence level.When confidence level is less than certain threshold value, it was demonstrated that tracking object is blocked, and runs YOLO detection process at this time.YOLO detects multiple alternative targets, makees essence detection, the maximum as tracking target of the value that meets with a response according to the tracking filter model and alternative target that have learnt at this time.This method tracking target is accurate, and real-time is good.
Description
Technical field
The present invention relates to object tracking methods more particularly to a kind of object tracking methods of view-based access control model.
Background technology
Tracking moving object is one of research emphasis problem of computer vision field, is all had in many occasions important
Using, with reference to three-axis stabilization holder realize tracing control function be one of its application scenario.The speed of service is fast at present and effect compared with
Good tracking is that nuclear phase closes algorithm filter KCF and is with the addition of the DSST algorithms of dimension self-adaption function.KCF is used
The circular matrix of target peripheral region gathers positive negative sample, using ridge regression training objective detector, using circular matrix in Fu
In vane space can the property of diagonal angling the computing of matrix is converted into point multiplication operation, greatly reduce operand, improve fortune
Speed is calculated, algorithm is made to meet requirement of real-time.
KCF track algorithms have a defect that model drifts about at any time, it is impossible to well adapt to as blocked, illumination variation
Greatly, situations such as background interference is larger can only be used, it is necessary to reference to detection module as short time tracking filter, lost in tracking
The rectangle frame of tracking module can be reset when effect.Detection process is generally time-consuming longer, because to be search for whole pictures
Processing reason.The method generally used is tracking-detection scheme as TLD algorithms, and individually one grader of training is used to examine
It surveys, the confidence level during correlation filter tracking, such as maximum response, peak sidelobe ratio etc. is calculated in real time, in high confidence level
Shi Gengxin tracking results and correlation filter model, while update the positive negative sample of detection module, during low confidence, operation detection
Module continues to track after detecting target.
The shortcomings that above-mentioned way is:1st, together with the update of model modification process and tracking result is, it is impossible to ensure well
Model drift can be reduced and keep up with target as far as possible;It, individually one detection point of training simultaneously when the 2nd, tracking process operation
Class device adds more computation burden, declines algorithm real-time;3rd, detection process use sliding window method, if inefficiency and
Trace model has drifted about, and is readily detected false target and tracks.
YOLO (You only look once) detection method is a kind of object detector end to end, passes through point in space
The probability of bounding box and respective classes is cut into, using individual network directly from the general of whole image predicted boundary frame and classification
Rate.By the training pattern on Pascal VOC (a kind of to assess image classification, detection, the data set split), 20 are can detect that
Remaining kind of common object, such as people, animal, the vehicles, indoor object.
Better familiar object tracking can be realized with reference to YOLO detectors and KCF tracking filters.With prior art phase
Than, it is an advantage of the invention that:High confidence level updates tracking filter and greatly reduces filter drift;It is not required to during tracking
Individual grader is trained to improve real-time;Operation deep learning neutral net does object detection, obtains more accurately examining
It surveys as a result, being influenced smaller by trace model drift.
The content of the invention
In view of the above problems, for overcome the deficiencies in the prior art, the present invention provides a kind of tracking better base of accuracy
In the object tracking methods of vision.
A kind of object tracking methods of view-based access control model, include the following steps:
S1:Subscriber frame runs tracking process after selecting target to be tracked, calculates the peak sidelobe ratio during tracking in real time
PSR when PSR is more than threshold tau 1, starts KCF tracking and KCF detection process, updates tracking result and correlation filter model,
When PSR is between τ 1 and τ 2, tracking result is only updated, does not update correlation filter model, when PSR is less than τ 2, do not performed
Tracking process runs detection module;
S2:, tracking during train individual grader be used for detection module;
S3:YOLO detectors are run during detection, detect in image institute it is possible that comprising 20 kinds of common objects as into
The input of one step detection;
S4:Testing result and tracking filter model are made into multiplication cross, when result is more than threshold value P, update tracking result, and
Detector out of service.
Further, the KCF tracking process includes the following steps:
S11, in It frames, sampled near the pt of current location, training one recurrence device, it is described recurrence device for calculate one
The response of a wicket sampling;
S12, in It+1 frames, sampled near previous frame position pt, judge the sound each sampled with the recurrence device in S11
It should;
The most strong sampling of S13, response is as this frame position pt+1.
Further, the operation YOLO detectors in the S3, specifically comprise the following steps:
When S31, tracking failure, YOLO target detection processes are activated;
Output after S32, detection is the result is that output boundary frame, the XY coordinates including bounding box rectangle, wide height, the value of the confidence,
And generic etc..
S33, according to testing result, obtains several roughing results;
S34, roughing result is carried out selected, obtains unique maximum response, be final detection target.
Further, the object tracking methods are on Macbook Air 1.7GHz i7 processors, using Cmake+
OpenCV development platforms emulate OTB100 sequences.
Further, the object tracking methods are being based on Xcode+OpenCV development platforms, trace routine on mobile phone
Real-time Transmission target frame position, from steady holder, realizes the tracing control function to target to three axis.
Further, the object tracking methods can be used for tracking a variety of different types of objects.
Further, the object tracking methods can be used for the object of 20 kinds or more of tracking.
Further, the object tracking methods are applied on mobile phone.
Further, the object tracking methods are applied on the mobile phone of hand-held holder.
In conclusion the present invention relates to a kind of object tracking methods of view-based access control model, tracked with reference to YOLO detectors and KCF
Wave filter can realize better familiar object tracking.Compared with prior art, it is an advantage of the invention that:High confidence level updates
Tracking filter greatly reduces filter drift;Individual grader need not be trained to improve real-time during tracking;Fortune
Row deep learning neutral net does object detection, obtains more accurate testing result, is influenced smaller by trace model drift.
Description of the drawings
With reference to appended attached drawing, more fully to describe the embodiment of the present invention.However, appended attached drawing be merely to illustrate and
It illustrates, and is not meant to limit the scope of the invention.
Fig. 1 is the object tracking flow chart of the prior art.
Fig. 2 is a kind of object tracking flow chart of view-based access control model of the present invention.
Specific embodiment
Below in conjunction with the accompanying drawings 1 and attached drawing 2 the present invention is further illustrated, it is but not as limiting to the invention.
KCF tracking process is as follows:
According to current frame information and before, frame information trains a correlation filter to correlation filtering, is carried out with the frame newly inputted
Correlation calculations, obtained response diagram are exactly the tracking result predicted, equation below 1:
F (z)=<W, z>
Wherein w represents correlation filter model parameter, and z represents the image block where target.
KCF track the step of be:
In It frames, (adding padding) is sampled near the pt of current location, one recurrence device of training.This returns device energy
Calculate the response of a wicket sampling;
It in It+1 frames, is sampled near previous frame position pt, judges the response each sampled with foregoing recurrence device;
The most strong sampling of response is as this frame position pt+1.
Using initial frame image block is selected to carry out correlation filter model training first, trained target is to minimize hits
According to the distance of the true tag yi (regressive object) of the calculating label f (xi) and next frame locations of real targets of xi.It is returned using ridge
Return, process equation below 2:
xm,nRepresent each image block for training pattern, i.e. original sample collection, it is added by initial tracking box
Padding and Cosine Window carry out cyclic shift and obtain again.λ is regularization parameter, ym,nIt is trained regressand value f (xm,n), i.e., it is original
The position of corresponding Gauss map after sample loops displacement.
Using cyclic shift characteristic, accelerate to solve in frequency domain, as shown in formula 3:
Represent the Fourier transformation of w.
Above-mentioned is that linear model solves, and using kernel function, actual nonlinear model is transformed into linear model, obtains model
Shown in renewal process equation below 4:
Wherein ∧ represents Fourier transformation,For the nonlinear correlation filter model parameters in frequency domain,For training sample
This auto-correlation kernel function.
Shown in the quick detection process equation below 5 of target:
Z represents candidate detection block,Represent the kernel function between all training samples and candidate detection block, f represents frequency domain
Correlation filtering response namely z all cyclic shifts response, by f carry out IFFT obtain the correlation filtering of time domain
Response, maximum value position represent the position where tracking target.
Tracking creditability is calculated, according to peak sidelobe ratio, equation below 6:
Wherein fmaxRepresent correlation filtering response maximum response max, definition secondary lobe is N × N image blocks, and μ and σ divide
Not Biao Shi secondary lobe average value and variance.
Using τ, threshold tau 1 and τ 2 are defined.
As T < τ 1, tracking failure stops tracking, is transferred to detection process;
It as 1 < T < τ 2 of τ, is updated without filter model, only updates tracking result;
As T > τ 2, while update tracking result and tracking filter.
Detection process is as follows:
When tracking failure, YOLO target detection processes are activated.The characteristics of YOLO is maximum is that the speed of service is fast,
Detection speed on iphone6s can reach 8fps, meet the requirement of real-time to detection.Pass through training on Pascal VOC
Model, detection more than 20 kinds of different class of identification, such as:Automobile, bicycle, ship, animal, people etc..
Output after detection is the result is that output boundary frame, the XY coordinates including bounding box rectangle, wide height, the value of the confidence and
Generic etc..
According to testing result, obtain being more than the roughing within about 10 as a result, to carry out essence to roughing result in next step
Choosing, obtains unique maximum response, is final detection target.
If the classification of tracking target has been preset, such as setting tracking pedestrians, then roughing result can be further
It reduces, and then improves detection efficiency.Further, if pedestrian's quantity is unique in setting tracking environmental, can skip over selected
Process as long as target is reappeared in angular field of view and detected by YOLO detectors, switches to tracking process immediately.Therefore
And the method is applicable to different use occasions.
Smart detection process is as follows:
For above-mentioned YOLO bounding box P, because it may be any size, it is not necessarily initial frame and selects target sizes w1
×h1Integral multiple, so by its resize to sdw1×sdh1, wherein sdFor the padding values in KCF algorithms.
After resize pictures, trained correlation filter model before is utilizedIt is obtained using Fourier inversion
Model in Time Domain α.Target detection process be correlation filter model with candidate detection block when domain operation, as shown in formula 7:
Above formula represents the response maximum of each image block of calculating, then the extreme value in maximizing.Because need not pair
YOLO bounding boxes carry out cyclic shift, so need not be in frequency-domain calculations.The maximum f of f (p)max。
Detailed process is:
By image block p (YOLO bounding boxes) resize to s of candidatedw1×sdh1Size obtains new image block pr, sdw1
×sdh1Add padding for initial tracking box.
Obtain prHOG features pzf;
Fourier transformation is made to feature, is obtained
The kernel function of time domain is calculated
The response f (p) of time domain is calculated, obtains the corresponding maximum f ' of this bounding boxmax;
Above calculating is repeated, until having traveled through all YOLO bounding boxes, obtains wherein maximum fmax。
If threshold value fδ, work as fmax>fδWhen, it was demonstrated that current detection output valve is high with correlation filter Model Matching precision,
Corresponding bounding box is exported as detection, is inputted as KCF.
In general, it is as follows to track overall procedure:
1st, initialize
1) centered on frame selects target, it is target sizes to gather a size×The initial sample sz of padding sizes;
2) the Gauss value yf of regression value, i.e. sample image block is calculated, calculates Cosine Window cos_window;
3) feature is extracted to initial training sample sz, carries out cyclic shift after adding Cosine Window, that is, transform to frequency domain, obtain
xf;
4) kernel function is carried out to the value xf after initial training sample shift and calculates kf, filter model is obtained according to yf and kf
model_alphaf;
2nd, process is tracked
1) the candidate image block of input picture is obtained, by its resize initial sample size;
2) to the feature of candidate image block, cyclic shift is carried out after adding Cosine Window, that is, is transformed into frequency domain, obtains zf;
3) the value zf after being shifted to candidate samples carries out kernel function and calculates kzf, is obtained according to model model_alphaf and kzf
Accordingly scheme to correlation filter;
4) maximum point response_max in corresponding figure is calculated, and calculates peak sidelobe ratio psr;
5) when psr is more than F2, response_max positions are updated, and obtain training image blocks, by training image
Block extracts feature, and cyclic shift is carried out after adding cosine frame, obtains xf, calculates kf and obtains more new model model_alphaf;
6) when psr is more than F1 and less than F2, response_max positions are updated, but without model modification;
7) when psr is less than F1, tracking is stopped, into detection process;
3rd, detection process
1) bounding box is detected by YOLO in real time, and provides size and confidence level;
2) when confidence level is more than threshold value, smart detection is carried out, obtains candidate samples collection z_yolo;
3) Fourier inversion is carried out to correlation filter model model_alphaf, obtains model_alpha;
4) by the size resize of all candidate samples collection z_yolo elements to initial target size×Padding, as time
Sampling sheet;
5) feature is extracted to candidate samples, Cosine Window is added to carry out circular matrix, that is, is transformed into frequency domain, obtains zf_yolo;
6) gaussian kernel function kf_yolo is calculated by zf_yolo and frequency-domain model model_alphaf, and kernel function is converted
K_yolo is obtained to time domain;
7) correlation filtering product is made by time domain kernel function k_yolo and Model in Time Domain model_alpha, obtains time domain response
Figure calculates wherein maximum, obtains detection_max;
8) candidate image blocks of the detection_max as next step tracking, continues to track.
Above-mentioned tracing detection process relies primarily on following platform to realize, in Macbook Air1.7GHz i7 processors
On, using Cmake+OpenCV development platforms, OTB100 sequences are emulated, track algorithm is used based on dimensional variation
KCF algorithms, feature selecting HOG, YOLO detection program are the model that Tiny-YOLO is trained based on Pascal VOC.
Above-mentioned algorithm is realized on iphone6s, based on Xcode+OpenCV development platforms, trace routine real-time Transmission
Target frame position, from steady holder, realizes the tracing control function to target to three axis.
4th, simulation result
8 groups of more typical scenes of emulation testing on OTB100, it is contemplated that object rapid deformation blocks, goes out image
Deng the situation for influencing long-time stable tracking success rate, the results are shown in table below:
Three kinds of algorithms are compared, colouring discrimination is as follows:
KCF | Learning | High confidence level updates | Detection algorithm | |
It is red | \ | √ | × | Sliding window method |
Yellow | \ | √ | \ | Sliding window method |
Green | \ | × | \ | YOLO+ essence detections |
For two kinds of situations that typical target is disturbed, since the peak side-lobe being calculated compares very little, and it is red
The threshold value of algorithm is larger, so tracking failure;Yellow algorithm lacks drift as a result of high confidence level more new strategy making model
In the case of as far as possible tracking gone up target, tracking process does not fail.The 60th frame Fig. 1), after target reappears, green
Algorithm has obtained new tracking position of object by YOLO+ essence detections, and yellow algorithm detects target using sliding window method, red
Color method is since the target positive sample that front is trained is less, so not detecting new target location.
For three kinds of excessive situations of typical target deformation, since deformation is excessive, so red algorithm already tracks failure
, and although yellow algorithm employs high confidence level track algorithm, but is still unable to effective district partial objectives for and background, cause with
Track model drifts about, and has kept up with the target of mistake, and in tracking failure, and also do not detect target again to the last afterwards.
Green method is also tracked in former frames to fail, but detects target again by YOLO detectors, and increases afterwards
The training sample of trace model is added so that track always afterwards successfully.Demonstrate YOLO detectors for track add Shandong
Stick.
For three kinds of few situations of common training sample early period, model modification is caused not catch up with the variation of environment, interference
Situation.At this moment using high confidence level update can improve tracking success rate to a certain extent, red algorithm failure after, yellow and
Although green algorithm drifts about, still tracking.But since target deformation is happened at the former frames of video, Learning mistakes
The model sample of Cheng Xunlian is very little, causes, once tracking fails, target just again to be can't detect using sliding window method.Using YOLO
Detector approach is based on training pattern, less can be influenced by such case.
In conclusion the present invention relates to a kind of object tracking methods of view-based access control model, tracked with reference to YOLO detectors and KCF
Wave filter can realize better familiar object tracking.Compared with prior art, it is an advantage of the invention that:High confidence level updates
Tracking filter greatly reduces filter drift;Individual grader need not be trained to improve real-time during tracking;Fortune
Row deep learning neutral net does object detection, obtains more accurate testing result, is influenced smaller by trace model drift,
Can improve well run into when tracking for a long time block, disturb, training sample early period few, situations such as target quickly changes,
Improve the robustness of tracking for a long time.
By explanation and attached drawing, the exemplary embodiments of the specific structure of specific embodiment are given, it is smart based on the present invention
God can also make other conversions.Although foregoing invention proposes existing preferred embodiment, however, these contents are not intended as
Limitation.
For a person skilled in the art, after reading above description, various changes and modifications undoubtedly will be evident.
Therefore, appended claims should regard whole variations and modifications of the true intention and scope that cover the present invention as.It is weighing
The scope and content of any and all equivalence, are all considered as still belonging to the intent and scope of the invention in the range of sharp claim.
Claims (9)
1. a kind of object tracking methods of view-based access control model, which is characterized in that include the following steps:
S1:Subscriber frame runs tracking process after selecting target to be tracked, calculates the peak sidelobe ratio PSR during tracking in real time, when
When PSR is more than threshold tau 1, starts KCF tracking and KCF detection process, tracking result and correlation filter model are updated, when PSR exists
When between τ 1 and τ 2, tracking result is only updated, does not update correlation filter model, when PSR is less than τ 2, do not performed and tracked
Journey runs detection module;
S2:, tracking during train individual grader be used for detection module;
S3:YOLO detectors are run during detection, detect in image institute it is possible that comprising 20 kinds of common objects as further
The input of detection;
S4:Testing result and tracking filter model are made into multiplication cross, when result is more than threshold value P, tracking result is updated, and stops
Operation detector.
2. the object tracking methods of a kind of view-based access control model according to claim 1, which is characterized in that the KCF was tracked
Journey includes the following steps:
S11, in It frames, sampled near the pt of current location, training one recurrence device, it is described recurrence device for calculate it is one small
The response of window sample;
S12, in It+1 frames, sampled near previous frame position pt, judge the response each sampled with the recurrence device in S11;
The most strong sampling of S13, response is as this frame position pt+1.
A kind of 3. object tracking methods of view-based access control model according to claim 1, which is characterized in that the operation in the S3
YOLO detectors, specifically comprise the following steps:
When S31, tracking failure, YOLO target detection processes are activated;
Output after S32, detection is the result is that output boundary frame, the XY coordinates including bounding box rectangle, wide height, the value of the confidence and
Generic etc..
S33, according to testing result, obtains several roughing results;
S34, roughing result is carried out selected, obtains unique maximum response, be final detection target.
A kind of 4. object tracking methods of view-based access control model according to any one of claim 1 to 3, which is characterized in that institute
Object tracking methods are stated on Macbook Air 1.7GHz i7 processors, it is right using Cmake+OpenCV development platforms
OTB100 sequences are emulated.
A kind of 5. object tracking methods of view-based access control model according to any one of claim 1 to 3, which is characterized in that institute
It states object tracking methods and Xcode+OpenCV development platforms is being based on mobile phone, trace routine real-time Transmission target frame position is given
Three axis realize the tracing control function to target from steady holder.
A kind of 6. object tracking methods of view-based access control model according to any one of claim 1 to 3, which is characterized in that institute
Object tracking methods are stated to can be used for tracking a variety of different types of objects.
A kind of 7. object tracking methods of view-based access control model according to claim 6, which is characterized in that the object tracking side
Method can be used for the object of 20 kinds or more of tracking.
A kind of 8. object tracking methods of view-based access control model according to any one of claim 1 to 3, which is characterized in that institute
Object tracking methods are stated to apply on mobile phone.
A kind of 9. object tracking methods of view-based access control model according to claim 8, which is characterized in that the object tracking side
Method is applied on the mobile phone of hand-held holder.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711243635.4A CN108062764A (en) | 2017-11-30 | 2017-11-30 | A kind of object tracking methods of view-based access control model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711243635.4A CN108062764A (en) | 2017-11-30 | 2017-11-30 | A kind of object tracking methods of view-based access control model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108062764A true CN108062764A (en) | 2018-05-22 |
Family
ID=62136225
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711243635.4A Pending CN108062764A (en) | 2017-11-30 | 2017-11-30 | A kind of object tracking methods of view-based access control model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108062764A (en) |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108810616A (en) * | 2018-05-31 | 2018-11-13 | 广州虎牙信息科技有限公司 | Object localization method, image display method, device, equipment and storage medium |
CN108960286A (en) * | 2018-06-01 | 2018-12-07 | 深圳市茁壮网络股份有限公司 | A kind of target following localization method and device |
CN109063594A (en) * | 2018-07-13 | 2018-12-21 | 吉林大学 | Remote sensing images fast target detection method based on YOLOv2 |
CN109285179A (en) * | 2018-07-26 | 2019-01-29 | 昆明理工大学 | A kind of motion target tracking method based on multi-feature fusion |
CN109407697A (en) * | 2018-09-20 | 2019-03-01 | 北京机械设备研究所 | A kind of unmanned plane pursuit movement goal systems and method based on binocular distance measurement |
CN109584271A (en) * | 2018-11-15 | 2019-04-05 | 西北工业大学 | High speed correlation filtering tracking based on high confidence level more new strategy |
CN109670410A (en) * | 2018-11-29 | 2019-04-23 | 昆明理工大学 | A kind of fusion based on multiple features it is long when motion target tracking method |
CN109740654A (en) * | 2018-12-26 | 2019-05-10 | 华东师范大学 | A kind of tongue body automatic testing method based on deep learning |
CN109816693A (en) * | 2019-01-28 | 2019-05-28 | 中国地质大学(武汉) | Anti- based on multimodal response blocks correlation filtering tracking and systems/devices |
CN109859243A (en) * | 2019-01-18 | 2019-06-07 | 昆明理工大学 | A kind of motion target tracking method based on dimension self-adaption block particle |
CN110009060A (en) * | 2019-04-17 | 2019-07-12 | 东北大学 | A kind of robustness long-term follow method based on correlation filtering and target detection |
CN110211152A (en) * | 2019-05-14 | 2019-09-06 | 华中科技大学 | A kind of endoscopic instrument tracking based on machine vision |
CN110246155A (en) * | 2019-05-17 | 2019-09-17 | 华中科技大学 | One kind being based on the alternate anti-shelter target tracking of model and system |
CN110728694A (en) * | 2019-10-10 | 2020-01-24 | 北京工业大学 | Long-term visual target tracking method based on continuous learning |
CN110991272A (en) * | 2019-11-18 | 2020-04-10 | 东北大学 | Multi-target vehicle track identification method based on video tracking |
CN111402295A (en) * | 2020-03-11 | 2020-07-10 | 桂林理工大学 | Moving object identification method based on object detection and tracking |
WO2020187095A1 (en) * | 2019-03-20 | 2020-09-24 | 深圳市道通智能航空技术有限公司 | Target tracking method and apparatus, and unmanned aerial vehicle |
CN112184770A (en) * | 2020-09-28 | 2021-01-05 | 中国电子科技集团公司第五十四研究所 | Target tracking method based on YOLOv3 and improved KCF |
CN114708300A (en) * | 2022-03-02 | 2022-07-05 | 北京理工大学 | Anti-blocking self-adaptive target tracking method and system |
-
2017
- 2017-11-30 CN CN201711243635.4A patent/CN108062764A/en active Pending
Cited By (26)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108810616A (en) * | 2018-05-31 | 2018-11-13 | 广州虎牙信息科技有限公司 | Object localization method, image display method, device, equipment and storage medium |
US11284128B2 (en) | 2018-05-31 | 2022-03-22 | Guangzhou Huya Information Technology Co., Ltd. | Object positioning method, video display method, apparatus, device, and storage medium |
CN108960286A (en) * | 2018-06-01 | 2018-12-07 | 深圳市茁壮网络股份有限公司 | A kind of target following localization method and device |
CN109063594A (en) * | 2018-07-13 | 2018-12-21 | 吉林大学 | Remote sensing images fast target detection method based on YOLOv2 |
CN109285179B (en) * | 2018-07-26 | 2021-05-14 | 昆明理工大学 | Moving target tracking method based on multi-feature fusion |
CN109285179A (en) * | 2018-07-26 | 2019-01-29 | 昆明理工大学 | A kind of motion target tracking method based on multi-feature fusion |
CN109407697A (en) * | 2018-09-20 | 2019-03-01 | 北京机械设备研究所 | A kind of unmanned plane pursuit movement goal systems and method based on binocular distance measurement |
CN109584271A (en) * | 2018-11-15 | 2019-04-05 | 西北工业大学 | High speed correlation filtering tracking based on high confidence level more new strategy |
CN109670410A (en) * | 2018-11-29 | 2019-04-23 | 昆明理工大学 | A kind of fusion based on multiple features it is long when motion target tracking method |
CN109740654A (en) * | 2018-12-26 | 2019-05-10 | 华东师范大学 | A kind of tongue body automatic testing method based on deep learning |
CN109859243A (en) * | 2019-01-18 | 2019-06-07 | 昆明理工大学 | A kind of motion target tracking method based on dimension self-adaption block particle |
CN109816693A (en) * | 2019-01-28 | 2019-05-28 | 中国地质大学(武汉) | Anti- based on multimodal response blocks correlation filtering tracking and systems/devices |
WO2020187095A1 (en) * | 2019-03-20 | 2020-09-24 | 深圳市道通智能航空技术有限公司 | Target tracking method and apparatus, and unmanned aerial vehicle |
CN110009060A (en) * | 2019-04-17 | 2019-07-12 | 东北大学 | A kind of robustness long-term follow method based on correlation filtering and target detection |
CN110009060B (en) * | 2019-04-17 | 2021-07-23 | 东北大学 | Robustness long-term tracking method based on correlation filtering and target detection |
CN110211152A (en) * | 2019-05-14 | 2019-09-06 | 华中科技大学 | A kind of endoscopic instrument tracking based on machine vision |
CN110246155A (en) * | 2019-05-17 | 2019-09-17 | 华中科技大学 | One kind being based on the alternate anti-shelter target tracking of model and system |
CN110246155B (en) * | 2019-05-17 | 2021-05-18 | 华中科技大学 | Anti-occlusion target tracking method and system based on model alternation |
CN110728694A (en) * | 2019-10-10 | 2020-01-24 | 北京工业大学 | Long-term visual target tracking method based on continuous learning |
CN110728694B (en) * | 2019-10-10 | 2023-11-24 | 北京工业大学 | Long-time visual target tracking method based on continuous learning |
CN110991272A (en) * | 2019-11-18 | 2020-04-10 | 东北大学 | Multi-target vehicle track identification method based on video tracking |
CN110991272B (en) * | 2019-11-18 | 2023-07-18 | 东北大学 | Multi-target vehicle track recognition method based on video tracking |
CN111402295A (en) * | 2020-03-11 | 2020-07-10 | 桂林理工大学 | Moving object identification method based on object detection and tracking |
CN111402295B (en) * | 2020-03-11 | 2023-08-08 | 桂林理工大学 | Moving object identification method based on object detection and tracking |
CN112184770A (en) * | 2020-09-28 | 2021-01-05 | 中国电子科技集团公司第五十四研究所 | Target tracking method based on YOLOv3 and improved KCF |
CN114708300A (en) * | 2022-03-02 | 2022-07-05 | 北京理工大学 | Anti-blocking self-adaptive target tracking method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108062764A (en) | A kind of object tracking methods of view-based access control model | |
CN112132042B (en) | SAR image target detection method based on contrast domain adaptation | |
CN104200237B (en) | One kind being based on the High-Speed Automatic multi-object tracking method of coring correlation filtering | |
CN109325418A (en) | Based on pedestrian recognition method under the road traffic environment for improving YOLOv3 | |
Gao et al. | A novel target detection method for SAR images based on shadow proposal and saliency analysis | |
CN105894701B (en) | The identification alarm method of transmission line of electricity external force damage prevention Large Construction vehicle | |
CN109902715B (en) | Infrared dim target detection method based on context aggregation network | |
CN103729854B (en) | A kind of method for detecting infrared puniness target based on tensor model | |
CN106650600A (en) | Forest smoke and fire detection method based on video image analysis | |
CN105931269A (en) | Tracking method for target in video and tracking device thereof | |
CN104794729B (en) | SAR image change detection based on conspicuousness guiding | |
CN107507170A (en) | A kind of airfield runway crack detection method based on multi-scale image information fusion | |
CN106570893A (en) | Rapid stable visual tracking method based on correlation filtering | |
CN108647695A (en) | Soft image conspicuousness detection method based on covariance convolutional neural networks | |
CN103886325A (en) | Cyclic matrix video tracking method with partition | |
CN103336947A (en) | Method for identifying infrared movement small target based on significance and structure | |
CN110472607A (en) | A kind of ship tracking method and system | |
CN106897730A (en) | SAR target model recognition methods based on fusion classification information with locality preserving projections | |
CN110084834A (en) | A kind of method for tracking target based on quick tensor singular value decomposition Feature Dimension Reduction | |
CN114266805A (en) | Twin region suggestion network model for unmanned aerial vehicle target tracking | |
CN110188661A (en) | Boundary Recognition method and device | |
CN109727268A (en) | Method for tracking target, device, computer equipment and storage medium | |
CN104268852B (en) | The detection method and its detection means of Weak target in infrared sequence image | |
Li et al. | Multi-scale ships detection in high-resolution remote sensing image via saliency-based region convolutional neural network | |
Cui et al. | An infrared small target detection framework based on local contrast method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180522 |