CN109102522A - A kind of method for tracking target and device - Google Patents
A kind of method for tracking target and device Download PDFInfo
- Publication number
- CN109102522A CN109102522A CN201810768738.0A CN201810768738A CN109102522A CN 109102522 A CN109102522 A CN 109102522A CN 201810768738 A CN201810768738 A CN 201810768738A CN 109102522 A CN109102522 A CN 109102522A
- Authority
- CN
- China
- Prior art keywords
- filter
- frame
- target
- image
- response
- 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.)
- Granted
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
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- 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
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention proposes a kind of method for tracking target and device, this method comprises: initializing all filters according to given target position in initial frame image；Response computation is carried out for current frame image candidate frame matched filter.Peak value of response is greater than threshold value if it exists, then determines therefrom that target position；Peak value of response is greater than threshold value if it does not exist, then suspends filter update, into occlusion state；Frame, which is surrounded, according to previous frame target determines present frame object detection area.Candidate frame is extracted according to detection high score frame, matched filter and calculates peak value of response respectively, thinks that target is given for change if detection score and peak value of response are all larger than threshold value, is otherwise still occlusion state；Continuous multiple frames are blocked, and be will test range and are expanded to full figure, until giving target for change again.In tracking target carriage change, ambient lighting changes and all has stronger polymorphic adaptability and robustness situations such as blocking this method, is easy to implement all-the-way tracking and the observation of target.
Description
Technical field
The present invention relates to a kind of method for tracking target and devices, belong to technical field of image processing, are specifically exactly
A kind of the multi-filter target tenacious tracking algorithm and device of detection auxiliary.
Background technique
Target following has been widely used in computer vision, monitoring system, civilian safety check and infrared guidance etc. and has ground
Study carefully field.The essence of target following is position and geological information of the determining target in image sequence.Similar background interference and with
The difficulty for all accurately tracking long-time stable such as block of the similar object of track target greatly increases, it has also become computer vision neck
The research hotspot in domain.
The display model that method for tracking target uses is divided into two major classes: production model and discriminative model.Its is maximum
Difference is exactly that production model does not utilize background information and background information is utilized in discriminative model.I.e. production model utilizes
Each positive sample establishes the appearance data prior distribution of target, specific to determine target feature itself, and ignores background image information.
And discriminative model is also with the negative sample comprising background, positive negative sample can be separated very well and can be promoted by training
Classifier.Present track algorithm is mainly discriminative model, because background information can be made full use of.Correlation filter tracking is calculated
What method used is discriminative model,
Currently, having done many work for correlation filter class track algorithm, many effective improvement sides are proposed
Method, such as article " Henriques J F, Rui C, Martins P, et al, " High-Speed Tracking with
Kernelized Correlation Filters,”IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol.37, no.3, pp.583-596,2015. " propose to introduce the KCF tracking of gaussian kernel function
Algorithm, article " Danelljan M, G,Khan F S,“Accurate scale estimation for robust
visual tracking,”in British Machine Vision Conference,Nottingham,United
Kingdom, 2014, the dimensional variation of pp.65.1-65.11. " adaptation tracking target, article " Galoogahi H K, Sim T,
Lucey S,“Correlation filters with limited boundaries,”in IEEE Conference on
Computer Vision and Pattern Recognition, Boston, USA, 2015, pp.4630-4638. " reduction side
Boundary's effect, article " Danelljan M, Robinson A, Khan F S, et al, " Beyond Correlation
Filters:Learning Continuous Convolution Operators for Visual Tracking,”in
European Conference on Computer Vision,Amsterdam,The Netherlands,2016,pp.472-
488. " convert continuous characteristic image and the general continuous filter of training for discrete features image to merge different scale characteristic pattern
As information etc..Wherein, the KCF track algorithm of gaussian kernel function is introduced as a kind of track algorithm that performance is outstanding, in practical work
Relatively broad application has been obtained in journey.But it is blocked since KCF can not be identified, and nearest tracking mesh can only be stored
Logo image information, so target surrounds frame and is easy to happen drift when tracking target drastic mechanical deformation occurring, when target occlusion, will draw
Enter a large amount of error training samples, these all will lead to proving an abortion for tracking.
For tracking target drastic mechanical deformation and the problem of block, scholars are studied, as article " Ma C, Yang X,
Zhang C,et al,“Long-term correlation tracking,”in IEEE Conference on Computer
Vision and Pattern Recognition, Boston, USA, 2015, pp.5388-5396. " use are different with new rate
The regression model of sample copes with quick deformation and the conservative size estimation, article " Danelljan M, Hager of tracking target respectively
G,Khan F S,et al,“Adaptive Decontamination of the Training Set:A Unified
Formulation for Discriminative Visual Tracking,”in IEEE Conference on
Computer Vision and Pattern Recognition, Las Vegas, USA, 2016, pp.1430-1438. " study
The weight of training sample, article " Huang C, Wu B, Nevatia R, " Robust Object Tracking by
Hierarchical Association of Detection Responses,”in European Conference on
Computer Vision, Marseille, France, 2008, pp.788-801. " is tracked using pursuit path information auxiliary to be sentenced
It is disconnected block, article " Cehovin L, Kristan M, Leonardis A, " Robust visual tracking using an
adaptive coupled-layer visual mode,”IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol.35, no.4, pp.941-953,2013. ", which will be blocked, to be corresponded to target template and only has
It is the rarefaction representation etc. of nonzero element in image small cube position.Although these methods improve tracking to a certain extent and calculate
Method is for the robustness blocked, but rate request that is often computationally intensive, can not meet real-time tracking.In order to avoid these are asked
Topic, it may be considered that training sample is classified using the KCF track algorithm of multi-filter, a variety of tracking mesh not only can be recorded
Typical historical Appearance is marked, pollution of the error training sample to correct training sample can also be mitigated, in conjunction with the detection algorithm of introducing,
Identification is blocked, and realizes that target is given for change after target reappears, and is corrected target and surrounded frame position.
Summary of the invention
It is an object of the invention to propose a kind of method for tracking target and device, this method is in traditional correlation filter KCF
It on the basis of track algorithm, is improved using multi-filter structure, and introduces detection algorithm auxiliary tracking.
The technical solution adopted by the present invention are as follows: a kind of target tracker, by visible light optical focus switchable imaging system, control is calculated
Machine and two axis servo-systems composition, in which:
Visible light optical focus switchable imaging system is made of industrial camera and visible light zoom lens two parts.Industrial camera shoots mesh
Logo image, and image data is transferred to by control computer by converting transmission system, can darkening zoom lens then according to tracking
Computer feedback result controls lens focus, so that target size in shooting image is kept constant；
Control computer determines the specific location that target frame is tracked in present image using track algorithm, according to target following
Frame accounts for the big minor adjustment lens focus of image scaled, and the departure of the target position and field of view center point provided by track algorithm
Two axis servo-systems, the i.e. offset by target's center away from picture centre are controlled, to control two axis servo-systems accordingly make target
It is maintained at picture centre region.
Two axis servo-systems are mainly made of turntable stage body and electric cabinet two large divisions.Turntable stage body part be system most
Whole executing agency is respectively intended to complete the angular movement in orientation, pitching both direction using vertical U-shaped structure form.Control system
Positioned at mechanical stage body lower inside, for placing the control section of two axis servo-systems, receive the control letter of control computer
Number and realize the real time kinematics of each frame controlled.
A kind of method for tracking target realizes that steps are as follows:
Step 1: in initial frame image, clarification of objective image being extracted according to given target area, training obtains first
The characteristic image and weight coefficient of a filter, initialize other filters, successively perform the following operation:
It extracts target signature image: its corresponding hog characteristic image is extracted according to given target area；
Training weight coefficient: trained filter is that original training image cyclic shift obtains all images by non-linear
Transformed linear combination, the weight coefficient of linear combination are exactly training weight coefficient, calculation are as follows:
WhereinX is that c ties up hog characteristic image, and ^ represents direct computation of DFT
Leaf transformation, F-1Inverse discrete Fourier transform is represented,Same position element multiplication is represented,*Conjugation is represented, λ is regular coefficient, and σ is
Gaussian Profile mean square deviation, specific method is referring to " Henriques J F, Rui C, Martins P, et al, " High-Speed
Tracking with Kernelized Correlation Filters,”IEEE Transactions on Pattern
Analysis and Machine Intelligence,vol.37,no.3,pp.583-596,2015".The spy of remaining filter
It levies image and weight coefficient is identical as first filter, but respective filter weight is less than the filter of first filter
Weight.
Step 2；Candidate frame characteristic pattern is obtained with different scale in current frame image, respectively matched filter, if each
Filter fiCorresponding training characteristics image is Ifi.For current a certain scale candidate frame characteristic image I, in order to select best
With filter, calculates I and all N number of filters correspond to the similarity of training characteristics image, using following similarity evaluation side
Formula:
So that filter corresponding to the smallest characteristic image of S is most matched filter.
It is candidate frame characteristic image and matched filter convolution, the maximum of all positions of convolution results that peak value of response, which calculates,
Value is peak value of response.
Peak value of response is greater than threshold value if it exists, then chooses most matched filter and corresponding candidate frame according to peak response peak value
Characteristic image x calculates weight coefficient α using the candidate frame characteristic image chosen according to step 1 the method.Candidate frame characteristic pattern
As the filter information newly trained with gained weight coefficient i.e. representative is calculated, and the historical storage information of most matched filter is then used
Its corresponding history training characteristics image x ' and the α ' expression of history weight vectors.Most matched filter update mode are as follows: history instruction
Practice image x ' and be updated to θ x+ (1- θ) x ', history weight vectors α ' is updated to θ α+(1- θ) α '.
And increase the filter weight of most matched filter, accordingly reduce other filter weights:
Wherein fβIt is most matched filter,It is the corresponding filter weight of most matched filter, ωiIt is other filters
Corresponding filter weight.Peak value of response is greater than threshold value if it does not exist, then enters occlusion state.
Step 3: suspending filter update when occlusion state, frame position is surrounded according to k-1 frame target and determines target in kth frame
Detection zone.Based on detection high score frame location information, different scale candidate frame characteristic image is extracted, matched filter and is counted respectively
Peak value of response is calculated, thinks that target is given for change again if detection score and filter peak value of response are all larger than threshold value, jumps out and block shape
State, otherwise next frame is still directly entered occlusion state:
Object detection area is still that rectangle surrounds frame in kth frame, and center image coordinate and k-1 frame target rectangle surround frame
Center is identical, and long and wide respectively k-1 frame target rectangle surrounds the long and wide specific factor of frame.It is true according to hough transform high score frame
When determining candidate frame characteristic image, hough transform high score frame is identical with candidate frame center image coordinate, the length and width point of candidate rectangle frame
Not Wei hough transform frame length and width specific factor η, extract different scale candidate frame i.e. by η value be different numerical value carry out characteristic pattern
As extracting
Step 4: continuous 20 frame, which blocks and do not give target for change, then enters the state of emergency, extended detection range to full figure, i.e., not
It is only limitted to previous frame target and surrounds frame position and its peripheral region, but full figure is detected, remaining will test high score candidate frame
Matched filter simultaneously calculates the succeeding targets such as peak value of response to give step for change identical as the algorithm described in step 2.
Step 5: after giving target for change again, end is blocked or the state of emergency, if it exists the filter weight of some filter
The filter is then updated less than threshold value, otherwise updates most matched filter, and the filter weight of increase update filter is simultaneously corresponding
Reduce the filter weight of other filters.
Theoretical basis of the invention is the multi-filter track algorithm of detection auxiliary, and implementation method is step 1-5, complete
Algorithm realizes that block diagram is as shown in Figure 4.It is in place of main innovation of the invention by all history training images according to similarity point
Class is different training image collection, and training obtains multiple filters corresponding to different training image collections (i.e. different target historical Appearance)
Wave device, multiple filter collective effects determine the picture position of current tracking target, and weight coefficient is by currently tracking target shape
State and the matching degree of each filter determine that currently determining tracking target surrounds frame and is also served only for updating corresponding matching filter
Wave device.Another innovation of the invention is to introduce detection auxiliary tracking, carries out shadowing and target is given for change.
Advantages of the present invention and effect are: not only can remember a variety of historical Appearances and adapt to the discontinuous of tracking target
Display model can also separate error training sample with correct training sample, not pollute introducing error training sample
The historical storage information of other correct filters.It introduces detection to be corrected, it is ensured that the accuracy of tracking box.Detection combines more filters
The a variety of tracking target appearance information judgement of wave device storage is blocked the even similar object of target and is blocked, can be to avoid error sample
Introducing, can also during target is given for change again, reduce detection range (short time block only need to detect script target time
Select frame and its peripheral region), exclude the interference of the tracking similar object of target, give the mesh in any history representative configuration for change
Mark.
Detailed description of the invention
Fig. 1 is device detecting and tracking module flow diagram；
Fig. 2 is two axis servo-system schematic diagrames；
Fig. 3 is visible light optical focus switchable imaging system；
Fig. 4 is that complete algorithm realizes block diagram in the present invention；
Fig. 5 is the schematic diagram that target is lost and given for change again in the embodiment of the present invention；
Fig. 6 is aircraft tracking effect schematic diagram in the embodiment of the present invention.
Specific embodiment
Device controls computer and two axis servo-systems composition by visible light optical focus switchable imaging system, in which:
Visible light optical focus switchable imaging system is made of industrial camera and visible light zoom lens two parts.Industrial camera shoots mesh
Logo image, and image data is transferred to by control computer by converting transmission system, can darkening zoom lens then according to tracking
Computer feedback result controls lens focus, so that target size in shooting image is kept constant；
Control computer determines the specific location that target frame is tracked in present image using track algorithm, according to target following
Frame accounts for the big minor adjustment lens focus of image scaled, and the departure of the target position and field of view center point provided by track algorithm
Two axis servo-systems, the i.e. offset by target's center away from picture centre are controlled, to control two axis servo-systems accordingly make target
It is maintained at picture centre region.
Two axis servo-systems are mainly made of turntable stage body and electric cabinet two large divisions.Turntable stage body part be system most
Whole executing agency is respectively intended to complete the angular movement in orientation, pitching both direction using vertical U-shaped structure form.Control system
Positioned at mechanical stage body lower inside, for placing the control section of two axis servo-systems, receive the control letter of control computer
Number and realize the real time kinematics of each frame controlled.
Target tenacious tracking method realizes that steps are as follows:
Step 1: in initial frame image, clarification of objective image being extracted according to given target area, training obtains first
The characteristic image and weight coefficient of a filter, initialize other filters, successively perform the following operation:
It extracts target signature image: its corresponding hog characteristic image is extracted according to given target area；
Training weight coefficient: more to give full expression to characteristic image information and then improving tracking accuracy, KCF author will be original
The large-scale characteristics image that characteristic image nonlinear transformation is amplified to length and width, and the corresponding large scale filter of training.Large scale
Filter is that original training image cyclic shift obtains linear combination of all images after same nonlinear transformation.Large scale
Characteristic image and all corresponding position element products of large scale filter and (dot product) be that original image center point target is deposited
Score.Trained large scale filter is that original training image cyclic shift obtains all images after nonlinear transformation
Linear combination, the weight coefficient of linear combination be exactly training weight coefficient, calculation are as follows:
WhereinX is that c ties up hog characteristic image, and ^ represents direct computation of DFT
Leaf transformation, F-1Inverse discrete Fourier transform is represented,Same position element multiplication is represented,*Conjugation is represented, λ is regular coefficient, σ
It is Gaussian Profile mean square deviation, specific method is referring to " Henriques J F, Rui C, Martins P, et al, " High-Speed
Tracking with Kernelized Correlation Filters,”IEEE Transactions on Pattern
Analysis and Machine Intelligence,vol.37,no.3,pp.583-596,2015".The spy of remaining filter
It levies image and weight coefficient is identical as first filter, but respective filter weight is less than the filter of first filter
Weight.
Step 2；The position that frame is surrounded based on previous frame target obtains candidate frame spy in current frame image with different scale
Sign figure, difference matched filter, if each filter fiCorresponding training characteristics image is Ifi.It is candidate for current a certain scale
Frame characteristic image I calculates I and all N number of filters corresponds to the similar of training characteristics image to select best match filter
Degree, using following similarity evaluation mode:
So that filter corresponding to the smallest characteristic image of S is most matched filter.
It is candidate frame characteristic image and matched filter convolution, the maximum of all positions of convolution results that peak value of response, which calculates,
Value is peak value of response.
Peak value of response is greater than threshold value if it exists, then chooses most matched filter and corresponding candidate frame according to peak response peak value
Characteristic image x calculates weight coefficient α using the candidate frame characteristic image chosen according to step 1 the method.Candidate frame characteristic pattern
As the filter information newly trained with gained weight coefficient i.e. representative is calculated, and the historical storage information of most matched filter is then used
Its corresponding history training characteristics image x ' and the α ' expression of history weight vectors.Most matched filter update mode are as follows: history instruction
Practice image x ' and be updated to θ x+ (1- θ) x ', history weight vectors α ' is updated to θ α+(1- θ) α '.And increase most matched filter
Filter weight accordingly reduces other filter weights:
Wherein fβIt is most matched filter,It is the corresponding filter weight of most matched filter, ωiIt is other filters
Corresponding filter weight.Peak value of response is greater than threshold value if it does not exist, then enters occlusion state.
Step 3: suspending filter update when occlusion state, detection zone is calculated according to the position that previous frame target surrounds frame
Domain will test region and be set as M times that target surrounds frame, obtains detection zone image.Object detection area is still in present frame
Rectangle surrounds frame, and center image coordinate is identical as previous frame target rectangle encirclement frame center, and long and width is respectively previous frame mesh
It marks rectangle and surrounds the long and wide specific factor of frame.When determining candidate frame characteristic image according to hough transform high score frame, hough transform
High score frame is identical with candidate frame center image coordinate, and the length and width of candidate rectangle frame are respectively the specific factor of hough transform frame length and width
η, extracting different scale candidate frame is that different numerical value carry out characteristic image extraction by η value.SSD detects institute in detection zone
There is the tracking similar object of target, detection score is below threshold value and then thinks to block, and otherwise, obtains detection score d and is greater than threshold value ν 1
Detection block.In the case where target congener soma disturbs more, need to will test frame, matched filter carries out peak response meter again
It adds and confirms that give for change is the same target, based on detection high score frame location information, extract different scale candidate frame characteristic pattern
Picture matched filter and calculates peak value of response respectively, thinks mesh if detection score and filter peak value of response are all larger than threshold value
Indicated weight is newly given for change, jumps out occlusion state, otherwise next frame is still directly entered occlusion state.There is no congener soma disturb with
Under track environment, can omit calculate detection block again matched filter carry out peak response calculating the step of.
Step 4: continuous 20 frame, which blocks and do not give target for change, then enters the state of emergency, extended detection range to full figure, i.e., not
It is only limitted to previous frame target and surrounds frame position and its peripheral region, but full figure is detected, and will test the matching of high score candidate frame
Filter simultaneously calculates the succeeding targets such as peak value of response to give step for change identical as the algorithm described in step 2.
Step 5: after giving target for change again, end is blocked or the state of emergency, if it exists the filter weight of some filter
The filter is then updated less than threshold value, otherwise updates most matched filter, and the filter weight of increase update filter is simultaneously corresponding
Reduce the filter weight of other filters.
Embodiment
Technical solution of the present invention is described in further detail below by way of specific embodiment.
Device controls computer and two axis servo-systems composition by visible light optical focus switchable imaging system.Visible light varifocal imaging
System is made of industrial camera and visible light zoom lens two parts.Industrial camera photographic subjects image, and pass through converting transmission
Image data is transferred to control computer by system；
Control computer determines that target surrounds frame position in current frame image using track algorithm, and surrounds frame by target
The deviation of center and field of view center point controls two axis servo-systems, the i.e. offset by target's center away from picture centre, accordingly
Controlling two axis servo-systems makes target be maintained at picture centre region.Control computer also passes through target in current frame image and surrounds
Frame accounts for the scale control lens focus of whole image, so that target size in shooting image is kept constant.System
Detecting and tracking module flow diagram is as shown in Figure 1.
Two axis servo-systems complete respective two-dimensional angular movement under the instruction of setting, so that test uses.Turntable stage body portion
Point it is the final executing agency of system, using vertical U-shaped structure form, is respectively intended to complete the angle of orientation, pitching both direction
Movement.Control system is located at mechanical stage body lower inside, for placing the control section of two axis servo-systems, realizes to each frame
Real time kinematics control and manipulation, monitoring, protection etc. functions.Various control instructions are inputted by rocking bar, are realized each to turntable
Kind control function.As shown in Figure 2.
Visible light optical focus switchable imaging system is made of visible light zoom lens and industrial camera two parts.Visible light zoom lens
Successively it is made of five parts such as focusing component, zoom component, rear fixation kit, seal assembly, photomoduel.Industrial camera choosing
The SP-5000 model high clear colorful camera designed with Japanese JAI company, as shown in Figure 3.
It is proposed that algorithm and device, there are similar background interference, block even tracking target congener to one group using the invention
Pedestrian movement's image sequence that body blocks is tracked using track algorithm of the invention, image be size be 640 × 480, position
The jpg image of depth 24, image sequence totalframes are 3624 frames, track target pedestrian 9 times and are blocked, block totalframes and reach
1379 frames.Testing platform used is ubuntu16.04.All experiments are in configuration intel core i7 GPU, dominant frequency
2.81Ghz, 8GB memory, video card be 1050 Ti of NVIDIA GeForce GTX computer on complete.Other parameters are equal in experiment
Code default parameters is provided using original author.
Fig. 4 gives the system flow chart of the multi-filter tracking of detection auxiliary.Shelter target gives effect for change again
As shown in figure 5, target pedestrian of every behavior is blocked, determine to block, the process that target pedestrian reappears and gives for change.
In the actual environment that track algorithm uses, when the interference of the similar object of other targets is less, it is convenient to omit detection
High score candidate frame matched filter calculates the step of filter response, but directly uses testing result as target position.It is real
The ECO having chosen using depth characteristic is tested, the KCF of hog feature and proposed in this paper improved by KCF has been used to detect the more of auxiliary
To 10 sections, there is no block interference but target morphology and change apparent aircraft video to be tracked jointly for filter tracks method.
Tracking process screenshot is as shown in fig. 6, every row is followed successively by ECO, KCF and context of methods tracking box during tracking from left to right
Position view.ECO specific method is referring to " Danelljan M, Bhat G, Khan F S, et al, " ECO:Efficient
Convolution Operators for Tracking, " in IEEE Conference on Computer Vision and
Pattern Recognition, State of Hawaii, USA, 2017, pp.6931-6939 ", KCF specific method referring to
" Henriques J F, Rui C, Martins P, et al, " High-Speed Tracking with Kernelized
Correlation Filters, " IEEE Transactions on PatternAnalysis and Machine
Intelligence, vol.37, no.3, pp.583-596,2015 ".
Although obvious tracking effect is more unstable, when aircraft does not take off also, due to scene it can be seen that KCF speed is fast
Air stream turbulence is obvious, and picture image quality is bad, and obviously deviating already occurs in the tracking box of KCF.In comparison ECO stablizes, but
After taking off, since apparent variation occurs in aircraft configuration, ECO can not also adapt to so violent deformation, tracking box
Also it has been gradually deviated from target.But method proposed in this paper from beginning to end can good tracking aircraft, there is no out
The drift of existing tracking box.It tests concrete outcome data and the comparison of several trackings is as shown in table 1, wherein FPS is in the track
The picture frame number exactly per second that can be tracked, this is related with the size and resolution ratio of picture, and average CLE is average central mistake
Difference is defined as the target's center of algorithm positioning and the average Euclidean distance at datum target center, and average OR is average coverage rate, fixed
Justice is the target's center of algorithm positioning and the average Euclidean distance at datum target center.
The comparison of 1 ECO, KCF and this paper algorithm keeps track all data of table
| Video name | Land1 | Land2 | Land3 | Land4 | Land5 | Launch1 | Launch2 | Launch3 | Launch4 | Launch5 | Average value |
| Length (frame) | 4800 | 7200 | 1997 | 2760 | 5101 | 4800 | 3600 | 3449 | 3960 | 4640 | 4230.7 |
| FPS(ECO) | < 10 | < 10 | < 10 | < 10 | < 10 | < 10 | < 10 | < 10 | < 10 | < 10 | < 10 |
| FPS(KCF) | 55 | 70 | 57 | 47 | 64 | 61 | 62 | 34 | 73 | 87.5 | 61.05 |
| FPS (improvement) | 48 | 34 | 69 | 74.6 | 62 | 33.6 | 50 | 34 | 43.5 | 45 | 49.37 |
| Average CLE (ECO) | 90.5 | 74.9 | 131.6 | 94.5 | 81 | 51.7 | 43 | 81.8 | 36 | 46.3 | 73.13 |
| Average CLE (KCF) | 55 | 37 | 94 | 48.5 | 45 | 60 | 74 | 105 | 30.5 | 43.8 | 59.28 |
| Average CLE (improvement) | 26.43 | 23 | 26 | 34.6 | 30.5 | 15| | 25 | 17.6 | 16.7 | 31.8 | 24.663 |
| Average OR (ECO) | 0.76 | 0.72 | 0.65 | 0.8 | 0.75 | 0.76 | 0.82 | 0.63 | 0.82 | 0.75 | 0.746 |
| Average OR (KCF) | 0.7 | 0.77 | 0.73 | 0.73 | 0.7 | 0.73 | 0.78 | 0.71 | 0.73 | 0.68 | 0.726 |
| Average OR (improvement) | 0.92 | 0.92 | 0.88 | 0.87 | 0.83 | 0.94 | 0.9 | 0.93 | 0.93 | 0.81 | 0.893 |
It is 1920*1080 pixel, the jpg image of bit depth 24 although 10 sections of videos include the in the same size of image.But
It is that the size specific gravity of target in the picture is different, target specific gravity has been largely fixed calculation amount and tracking velocity.Because
KCF surrounds block diagram picture training filter just for a frame target every time, and filter is after target surrounds block diagram as cyclic shift
Weight combination, and weight coefficient can directly be calculated, so its speed in most of video-frequency band is most fast.Real-time one
As require frame frequency more than 25 frames, it is more than substantially requirement of real time speed that KCF can satisfy even completely, but its precision obtains not
To guarantee.Average OR drifts about obvious always 0.7 or so in target appearance acute variation, and average CLE is in launch3 video
100 pixels have been even up in section.
ECO will surround block diagram picture in conjunction with the target of all historical storages every time and train continuous filter jointly, and can not
Filter is directly calculated, time-consuming conjugate gradient method need to be used to be iterated optimization, so it is all speed in all video-frequency bands
Most slow track algorithm, is not achieved requirement of real time completely.Average OR drifts about bright 0.75 or so in target appearance acute variation
Aobvious, average CLE has been even up to 130 pixels in land3.
The multi-filter innovatory algorithm velocity variations for introducing detection are larger, because it is detection and the combination of KCF, and detect
Speed with KCF be it is inconsistent, detection speed be slower than KCF speed, if in video-frequency band aircraft appearance acute variation number compared with
More, history representative configuration is more, need to repeatedly introduce detection, then time-consuming.Sometimes speed approaches even more than KCF to innovatory algorithm
The reason of be, KCF generate offset after, framed and much be not belonging to the background parts of target so increasing calculation amount, drag slowly
Speed, but innovatory algorithm more precisely frames target due to always being, so calculation amount is smaller, speed is promoted.Change
Frame frequency into algorithm is at least 30 or more, it is sufficient to meet requirement of real time, and innovatory algorithm precision is obviously improved, average OR is always
It is 0.8 or more, many times reaches 0.9, average CLE is also always within 35 pixels.
The above description is only an embodiment of the present invention, is not intended to limit protection scope of the present invention, all to utilize this hair
The protection of equivalent structure made by bright specification and accompanying drawing content or equivalent process transformation are applied directly or indirectly in other correlations
Technical field, similarly include within the scope of the present invention.
Claims (7)
1. a kind of target tracker, it is characterised in that: the device includes: visible light optical focus switchable imaging system, control computer, and two
Axis servo-system；Wherein,
Visible light optical focus switchable imaging system is made of industrial camera and visible light zoom lens two parts, industrial camera photographic subjects figure
Picture, and image data is transferred to by control computer by converting transmission system, can darkening zoom lens then calculated according to tracking
Machine feedback result controls lens focus, so that target size in shooting image is kept constant；
Control computer determines the specific location that target frame is tracked in present image using track algorithm, is accounted for according to target following frame
The big minor adjustment lens focus of image scaled, and the departure of the target position provided by track algorithm and field of view center point is controlled
Two axis servo-systems, the i.e. offset by target's center away from picture centre are made, controlling two axis servo-systems accordingly keeps target
Heart district domain in the picture；
Two axis servo-systems are mainly made of turntable stage body and electric cabinet two large divisions, and turntable stage body part is finally holding for system
Row mechanism is respectively intended to complete the angular movement in orientation, pitching both direction, control system is located at using vertical U-shaped structure form
Mechanical stage body lower inside receives the control signal of control computer simultaneously for placing the control section of two axis servo-systems
It realizes and the real time kinematics of each frame is controlled.
2. a kind of method for tracking target utilizes target tracker described in claim 1, it is characterised in that: realize step such as
Under:
Step 1: extracting clarification of objective image according to given target area, training obtains first filter in initial frame image
The characteristic image and weight coefficient of wave device, initialize other filters and its filter weight；
Step 2: obtaining candidate frame characteristic pattern in current frame image with different scale, matched filter carries out response peak respectively
Value calculates；Peak value of response is greater than threshold value if it exists, then chooses corresponding scale and filter according to peak response peak value, updates corresponding
Filter and increase its filter weight；Peak value of response is greater than threshold value if it does not exist, then enters occlusion state；
Step 3: suspending filter update when occlusion state, frame position is surrounded according to former frame target and determines target in present frame
Detection zone；Based on detection high score frame location information, different scale candidate frame characteristic image is extracted, matched filter and is counted respectively
Peak value of response is calculated, thinks that target is given for change again if detection score and peak value of response are all larger than threshold value, otherwise next frame is still direct
Into occlusion state；
Step 4: continuous 20 frame, which blocks and do not give target for change, then enters the state of emergency, extended detection range to full figure；
Step 5: end is blocked or the state of emergency, and the filter weight of some filter is less than if it exists after giving target for change again
Threshold value then updates the filter, otherwise updates most matched filter, increases the filter weight for updating filter and accordingly reduces
The filter weight of other filters.
3. method for tracking target according to claim 2, it is characterised in that:
In the step 1, in initial frame image, clarification of objective image is extracted according to given target area, training obtains first
The characteristic image and weight coefficient of a filter, initialize other filters, successively perform the following operation:
It extracts target signature image: its corresponding hog characteristic image is extracted according to given target area；
Training weight coefficient: trained filter is that original training image cyclic shift obtains all images by nonlinear transformation
Linear combination afterwards, the weight coefficient of linear combination are exactly training weight coefficient, the characteristic image and weight coefficient of remaining filter
It is identical as first filter, but respective filter weight is less than the filter weight of first filter.
4. method for tracking target according to claim 2, it is characterised in that:
In the step 2, candidate frame characteristic pattern is obtained with different scale in current frame image, respectively matched filter, if each
Filter fiCorresponding training characteristics image is Ifi；For current a certain scale candidate frame characteristic image I, in order to select best
With filter, calculates I and all N number of filters correspond to the similarity of training characteristics image, using following similarity evaluation side
Formula:
S=| | I-Ifi||2I=1,2...N
So that filter corresponding to the smallest characteristic image of S is most matched filter；
It is candidate frame characteristic image and matched filter convolution that peak value of response, which calculates, and the maximum value of all positions of convolution results is
For peak value of response；
Peak value of response is greater than threshold value if it exists, then chooses most matched filter and corresponding candidate frame feature according to peak response peak value
Image x calculates weight coefficient α using the candidate frame characteristic image chosen according to described；Candidate frame characteristic image and calculating gained
Weight coefficient is to represent newly trained filter information, and the historical storage information of most matched filter then uses its corresponding history
Training characteristics image x ' and the α ' expression of history weight vectors；Most matched filter update mode are as follows: history training image x ' update
θ α+(1- θ) α ' is updated to for θ x+ (1- θ) x ', history weight vectors α '；
And increase the filter weight of most matched filter, accordingly reduce other filter weights:
Wherein fβIt is most matched filter,It is the corresponding filter weight of most matched filter, ωiIt is that other filters are corresponding
Filter weight, if it does not exist peak value of response be greater than threshold value, then enter occlusion state.
5. method for tracking target according to claim 3, it is characterised in that:
In the step 3, when occlusion state, suspends filter update, surrounds frame position according to k-1 frame target and determines mesh in kth frame
Mark detection zone；Based on detection high score frame location information, different scale candidate frame characteristic image is extracted, difference matched filter is simultaneously
Peak value of response is calculated, thinks that target is given for change again if detection score and filter peak value of response are all larger than threshold value, jumps out and block
State, otherwise next frame is still directly entered occlusion state；
Object detection area is still that rectangle surrounds frame in kth frame, and center image coordinate and k-1 frame target rectangle surround frame center
Identical, long and wide respectively k-1 frame target rectangle surrounds the long and wide specific factor of frame；It is determined and is waited according to hough transform high score frame
When selecting frame characteristic image, hough transform high score frame is identical with candidate frame center image coordinate, and the length and width of candidate rectangle frame are respectively
The specific factor η of hough transform frame length and width, extracting different scale candidate frame is that different numerical value progress characteristic image mentions by η value
It takes.
6. method for tracking target according to claim 3, it is characterised in that:
In the step 4, continuous 20 frame, which blocks and do not give target for change, then enters the state of emergency, extended detection range to full figure, i.e., not
It is only limitted to previous frame target and surrounds frame position and its peripheral region, but full figure is detected, remaining will test high score candidate frame
Matched filter simultaneously calculates the succeeding targets such as peak value of response and gives for change.
7. method for tracking target according to claim 3, it is characterised in that:
In the step 5, after giving target for change again, end is blocked or the state of emergency, if it exists the filter weight of some filter
The filter is then updated less than threshold value, otherwise updates most matched filter, and the filter weight of increase update filter is simultaneously corresponding
Reduce the filter weight of other filters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810768738.0A CN109102522B (en) | 2018-07-13 | 2018-07-13 | Target tracking method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810768738.0A CN109102522B (en) | 2018-07-13 | 2018-07-13 | Target tracking method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109102522A true CN109102522A (en) | 2018-12-28 |
| CN109102522B CN109102522B (en) | 2021-08-31 |
Family
ID=64846336
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810768738.0A Active CN109102522B (en) | 2018-07-13 | 2018-07-13 | Target tracking method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109102522B (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109977928A (en) * | 2019-04-25 | 2019-07-05 | 中国科学院自动化研究所 | A kind of robot target pedestrian retrieval method |
| CN110009660A (en) * | 2019-03-06 | 2019-07-12 | 浙江大学 | Object space method for tracing based on correlation algorithm filter |
| CN110189365A (en) * | 2019-05-24 | 2019-08-30 | 上海交通大学 | It is anti-to block correlation filtering tracking |
| CN110210304A (en) * | 2019-04-29 | 2019-09-06 | 北京百度网讯科技有限公司 | Method and system for target detection and tracking |
| CN110290351A (en) * | 2019-06-26 | 2019-09-27 | 广东康云科技有限公司 | A kind of video target tracking method, system, device and storage medium |
| CN110400347A (en) * | 2019-06-25 | 2019-11-01 | 哈尔滨工程大学 | A kind of method for tracking target for judging to block and target relocates |
| CN110490907A (en) * | 2019-08-21 | 2019-11-22 | 上海无线电设备研究所 | Motion target tracking method based on multiple target feature and improvement correlation filter |
| CN110517483A (en) * | 2019-08-06 | 2019-11-29 | 杭州博信智联科技有限公司 | A kind of traffic information processing method and digital rail roadside unit |
| CN110599519A (en) * | 2019-08-27 | 2019-12-20 | 上海交通大学 | Anti-occlusion related filtering tracking method based on domain search strategy |
| CN110909604A (en) * | 2019-10-23 | 2020-03-24 | 深圳市华讯方舟太赫兹科技有限公司 | Security image detection method, terminal device and computer storage medium |
| CN112084914A (en) * | 2020-08-31 | 2020-12-15 | 的卢技术有限公司 | Multi-target tracking method integrating spatial motion and apparent feature learning |
| CN112585944A (en) * | 2020-01-21 | 2021-03-30 | 深圳市大疆创新科技有限公司 | Following method, movable platform, apparatus and storage medium |
| CN112862863A (en) * | 2021-03-04 | 2021-05-28 | 广东工业大学 | Target tracking and positioning method based on state machine |
| CN114842048A (en) * | 2022-04-11 | 2022-08-02 | 北京航天晨信科技有限责任公司 | Target tracking method, system, readable storage medium and computer equipment |
| US11748897B1 (en) * | 2022-06-24 | 2023-09-05 | Trans Astronautica Corporation | Optimized matched filter tracking of space objects |
| US12025006B2 (en) | 2020-09-22 | 2024-07-02 | Trans Astronautica Corporation | Systems and methods for radiant gas dynamic mining of permafrost |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651913A (en) * | 2016-11-29 | 2017-05-10 | 开易（北京）科技有限公司 | Target tracking method based on correlation filtering and color histogram statistics and ADAS (Advanced Driving Assistance System) |
| CN106686306A (en) * | 2016-12-22 | 2017-05-17 | 西安工业大学 | Target tracking device and target tracking method |
| CN107657630A (en) * | 2017-07-21 | 2018-02-02 | 南京邮电大学 | A kind of modified anti-shelter target tracking based on KCF |
| CN107748873A (en) * | 2017-10-31 | 2018-03-02 | 河北工业大学 | A kind of multimodal method for tracking target for merging background information |
| CN107993257A (en) * | 2017-12-28 | 2018-05-04 | 中国科学院西安光学精密机械研究所 | Intelligent IMM Kalman filtering feedforward compensation target tracking method and system |
| CN108010067A (en) * | 2017-12-25 | 2018-05-08 | 北京航空航天大学 | A kind of visual target tracking method based on combination determination strategy |
| CN108037510A (en) * | 2017-12-07 | 2018-05-15 | 武汉华之洋科技有限公司 | A kind of photoelectronic reconnaissance equipment for unmanned boat |
-
2018
- 2018-07-13 CN CN201810768738.0A patent/CN109102522B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651913A (en) * | 2016-11-29 | 2017-05-10 | 开易（北京）科技有限公司 | Target tracking method based on correlation filtering and color histogram statistics and ADAS (Advanced Driving Assistance System) |
| CN106686306A (en) * | 2016-12-22 | 2017-05-17 | 西安工业大学 | Target tracking device and target tracking method |
| CN107657630A (en) * | 2017-07-21 | 2018-02-02 | 南京邮电大学 | A kind of modified anti-shelter target tracking based on KCF |
| CN107748873A (en) * | 2017-10-31 | 2018-03-02 | 河北工业大学 | A kind of multimodal method for tracking target for merging background information |
| CN108037510A (en) * | 2017-12-07 | 2018-05-15 | 武汉华之洋科技有限公司 | A kind of photoelectronic reconnaissance equipment for unmanned boat |
| CN108010067A (en) * | 2017-12-25 | 2018-05-08 | 北京航空航天大学 | A kind of visual target tracking method based on combination determination strategy |
| CN107993257A (en) * | 2017-12-28 | 2018-05-04 | 中国科学院西安光学精密机械研究所 | Intelligent IMM Kalman filtering feedforward compensation target tracking method and system |
Non-Patent Citations (4)
| Title |
|---|
| JOA~O F. HENRIQUES, AT EL.: ""High-Speed Tracking with Kernelized Correlation Filters"", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| 江维创: ""基于核相关滤波器的目标跟踪方法研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 王鹏飞: ""基于似物性采样和核化相关滤波器的目标跟踪算法研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 马晓楠: ""电视导引头相关滤波跟踪算法研究"", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110009660A (en) * | 2019-03-06 | 2019-07-12 | 浙江大学 | Object space method for tracing based on correlation algorithm filter |
| CN109977928A (en) * | 2019-04-25 | 2019-07-05 | 中国科学院自动化研究所 | A kind of robot target pedestrian retrieval method |
| CN109977928B (en) * | 2019-04-25 | 2021-03-23 | 中国科学院自动化研究所 | Robot target pedestrian retrieval method |
| CN110210304A (en) * | 2019-04-29 | 2019-09-06 | 北京百度网讯科技有限公司 | Method and system for target detection and tracking |
| CN110189365A (en) * | 2019-05-24 | 2019-08-30 | 上海交通大学 | It is anti-to block correlation filtering tracking |
| CN110189365B (en) * | 2019-05-24 | 2023-04-07 | 上海交通大学 | Anti-occlusion correlation filtering tracking method |
| CN110400347A (en) * | 2019-06-25 | 2019-11-01 | 哈尔滨工程大学 | A kind of method for tracking target for judging to block and target relocates |
| CN110400347B (en) * | 2019-06-25 | 2022-10-28 | 哈尔滨工程大学 | Target tracking method for judging occlusion and target relocation |
| CN110290351B (en) * | 2019-06-26 | 2021-03-23 | 广东康云科技有限公司 | Video target tracking method, system, device and storage medium |
| CN110290351A (en) * | 2019-06-26 | 2019-09-27 | 广东康云科技有限公司 | A kind of video target tracking method, system, device and storage medium |
| CN110517483A (en) * | 2019-08-06 | 2019-11-29 | 杭州博信智联科技有限公司 | A kind of traffic information processing method and digital rail roadside unit |
| CN110490907A (en) * | 2019-08-21 | 2019-11-22 | 上海无线电设备研究所 | Motion target tracking method based on multiple target feature and improvement correlation filter |
| CN110490907B (en) * | 2019-08-21 | 2023-05-16 | 上海无线电设备研究所 | Moving target tracking method based on multi-target feature and improved correlation filter |
| CN110599519B (en) * | 2019-08-27 | 2022-11-08 | 上海交通大学 | Anti-occlusion related filtering tracking method based on domain search strategy |
| CN110599519A (en) * | 2019-08-27 | 2019-12-20 | 上海交通大学 | Anti-occlusion related filtering tracking method based on domain search strategy |
| CN110909604A (en) * | 2019-10-23 | 2020-03-24 | 深圳市华讯方舟太赫兹科技有限公司 | Security image detection method, terminal device and computer storage medium |
| CN110909604B (en) * | 2019-10-23 | 2024-04-19 | 深圳市重投华讯太赫兹科技有限公司 | Security check image detection method, terminal equipment and computer storage medium |
| CN112585944A (en) * | 2020-01-21 | 2021-03-30 | 深圳市大疆创新科技有限公司 | Following method, movable platform, apparatus and storage medium |
| CN112084914B (en) * | 2020-08-31 | 2024-04-26 | 的卢技术有限公司 | Multi-target tracking method integrating space motion and apparent feature learning |
| CN112084914A (en) * | 2020-08-31 | 2020-12-15 | 的卢技术有限公司 | Multi-target tracking method integrating spatial motion and apparent feature learning |
| US12025006B2 (en) | 2020-09-22 | 2024-07-02 | Trans Astronautica Corporation | Systems and methods for radiant gas dynamic mining of permafrost |
| CN112862863A (en) * | 2021-03-04 | 2021-05-28 | 广东工业大学 | Target tracking and positioning method based on state machine |
| CN112862863B (en) * | 2021-03-04 | 2023-01-31 | 广东工业大学 | Target tracking and positioning method based on state machine |
| CN114842048A (en) * | 2022-04-11 | 2022-08-02 | 北京航天晨信科技有限责任公司 | Target tracking method, system, readable storage medium and computer equipment |
| US11748897B1 (en) * | 2022-06-24 | 2023-09-05 | Trans Astronautica Corporation | Optimized matched filter tracking of space objects |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109102522B (en) | 2021-08-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109102522A (en) | A kind of method for tracking target and device | |
| US10782095B2 (en) | Automatic target point tracing method for electro-optical sighting system | |
| CN103455797B (en) | Detection and tracking method of moving small target in aerial shot video | |
| CN105373135B (en) | A kind of method and system of aircraft docking guidance and plane type recognition based on machine vision | |
| CN109816689A (en) | A kind of motion target tracking method that multilayer convolution feature adaptively merges | |
| CN109919974A (en) | Online multi-object tracking method based on the more candidate associations of R-FCN frame | |
| CN108109162B (en) | Multi-scale target tracking method using self-adaptive feature fusion | |
| CN104615986B (en) | The method that pedestrian detection is carried out to the video image of scene changes using multi-detector | |
| CN109977971A (en) | Dimension self-adaption Target Tracking System based on mean shift Yu core correlation filtering | |
| CN109741369A (en) | A kind of method and system for robotic tracking target pedestrian | |
| CN112734809B (en) | On-line multi-pedestrian tracking method and device based on Deep-Sort tracking framework | |
| CN109087337B (en) | Long-time target tracking method and system based on hierarchical convolution characteristics | |
| CN109099929A (en) | Intelligent vehicle positioning device and method based on scene fingerprint | |
| CN103677274A (en) | Interactive projection method and system based on active vision | |
| CN110245587B (en) | Optical remote sensing image target detection method based on Bayesian transfer learning | |
| CN109767454A (en) | Based on Space Time-frequency conspicuousness unmanned plane video moving object detection method | |
| CN107798691A (en) | A kind of unmanned plane independent landing terrestrial reference real-time detecting and tracking method of view-based access control model | |
| CN108846855A (en) | Method for tracking target and equipment | |
| CN111680713A (en) | Unmanned aerial vehicle ground target tracking and approaching method based on visual detection | |
| CN109448023A (en) | A kind of satellite video Small object method for real time tracking of combination space confidence map and track estimation | |
| CN110084830A (en) | A kind of detection of video frequency motion target and tracking | |
| CN114463619B (en) | Infrared dim target detection method based on integrated fusion features | |
| CN108986139B (en) | Feature integration method with significance map for target tracking | |
| CN103996207A (en) | Object tracking method | |
| CN106897730A (en) | SAR target model recognition methods based on fusion classification information with locality preserving projections |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |