CN108154159A - A kind of method for tracking target with automatic recovery ability based on Multistage Detector - Google Patents
A kind of method for tracking target with automatic recovery ability based on Multistage Detector Download PDFInfo
- Publication number
- CN108154159A CN108154159A CN201711422462.2A CN201711422462A CN108154159A CN 108154159 A CN108154159 A CN 108154159A CN 201711422462 A CN201711422462 A CN 201711422462A CN 108154159 A CN108154159 A CN 108154159A
- Authority
- CN
- China
- Prior art keywords
- target
- detector
- correlation
- degree
- sample
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- 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
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Probability & Statistics with Applications (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
Abstract
A kind of method for tracking target with automatic recovery ability based on Multistage Detector：The thought of strong classifier is cascaded as with reference to Adaboost multistage Weak Classifiers, the detector of multiple and different types is selected to connect；The first order detects and selects conspicuousness detector；Grader set detection module is selected in the second level, is updated to the posterior probability for calculating that the Probability Area that the first order detects is positive sample in random tree；The third level selects correlation filtering detector, the degree of correlation of positive sample that the sample of the big Mr. Yu's threshold value of posterior probability is obtained with initialization or previous frame is calculated, so as to reduce the accumulated error that long-term follow is brought；Determine that target area in present frame is regarded as, and sample with determining position in the position of degree of correlation maximum after Multistage Detector, additional sample concentrates the positive and negative sample number of removal, ensures that the reliability sum number purpose of sample is constant；If the small Mr. Yu's threshold value of degree of correlation maximum value, activates re-detection mechanism, search target is detected again to position near zone.
Description
Technical field
The present invention devises a kind of method for tracking target with automatic recovery ability based on Multistage Detector, by it is multiple not
Generic detector series connection, forming one strong detector, each frame image carries out target detection to video sequence, so as to fulfill
Target following, and re-detection can be carried out in the case where tracking target is lost, target location is detected again and continues to track.
This method all has preferable robust using the target following thinking based on detection for static background and dynamic background
Property, possess higher engineering application value.
Background technology
How maintain a long-term stability for image object the theoretical research of tracking, be increasingly becoming computer vision focus compared with
High direction has been widely used in the fields such as monitoring identification and robot.Target following is at a blending image signal
Many multi-disciplinary comprehensive study developing direction such as reason, pattern-recognition, automation, need to analyze and process image sequence,
So efficient, robust the tracker of design one is still a problem.Target following key problem in technology is to utilize feature or mould
Type detaches target with background difference, and common track algorithm is mainly the following thinking：The matched method of feature based, base
Method in Analysis of Contrast, the method based on model and the method for tracking target based on motion detection.Wherein based on movement mesh
The method main thought of mark detection is by the difference of the target and background in detection sequence image, determines region existing for target
With position so as to fulfill tracking.Traditional detection method has frame differential method, background subtraction and optical flow method etc..First two method
It is directly to do difference to front and rear two frame or background target area is calculated, this mode is only applicable to the processing of static background,
And it is easily influenced by moving non-targeted things in background.Although optical flow method can handle the problem of dynamic background,
Dense light stream point calculating needs to occupy a large amount of computing resource, and detection speed can not meet requirement of real-time.In addition with one
A little scholars propose new detection thinking, such as more hypothesis data correlation methods that Fu et al. is proposed, which is closed using probability data
The thought of connection solves the condition of uncertainty such as target loss to detecting target under complex scene.Ess etc. is using Kalman filtering as mesh
Target motion model.Data correlation is established using target appearance similarity, realizes the detection of target.Li et al., which proposes, utilizes statistics
Learn to solve the problems, such as that detection responds the thinking merged with target following.Wu proposes a sectional pattern to target, mesh
Probability-distribution function is marked as its feature, target is detected using Bayesian Fusion.But target is carried out using detection practical
During tracking, algorithm above common problem is the tracking error brought due to various possible factors, is being tracked
Error gradually accumulates in the process, and the detection aimed at precision after causing declines, and finally influences tracking effect.In addition target with
The initial phase of track, inherently there are accuracy problems for the target original template of acquisition, can also seriously affect and subsequently track
The precision of journey.If there are barrier or building in the track of target movement, target can due to being blocked features
Divide to disappear and even disappear completely, can not be detected again again when target occurs again and continue to track.
Invention content
The technology of the present invention solves the problems, such as：For in object tracking process by various environment influence and precision limited
Deficiency, a kind of method for tracking target with automatic recovery ability based on Multistage Detector is provided, improves entire tracking system
The robustness of system, and there is on-line study and removal accumulated error, enhance the engineering sense of tracking system.
The present invention technical solution be：It is a strong detector by several different types of weak detector series connection, carries
The accuracy of high target detection, and the method for combining on-line study effectively subtracts the positive negative sample real-time update in object module
The error that small long-time tracking accumulates in the process.In addition re-detection mechanism is established, again to target after target is in of short duration block
Region scans for detecting target again, further improves the tracking performance of the algorithm, and can successfully manage Various Complex ring
The interference of border factor.
A kind of method for tracking target with automatic recovery ability based on Multistage Detector of the present invention, including following step
Suddenly：First, the thought of strong classifier is cascaded as with reference to the more a Weak Classifiers of Adaboost, selects the detector of multiple and different types
Series connection, first order selection conspicuousness detector, calculating salient region of image, the position that quickly positioning target is likely to occur, second
Grade selection grader set detection module, is combined using multiple corner features, is updated in random tree and is calculated possible region and be
The posterior probability of positive sample, third level selection correlation filtering detector, calculate the sample of the big Mr. Yu's threshold value of posterior probability with it is initial
The degree of correlation of positive sample that change or previous frame obtain retains before relevancy ranking ten positive sample, it is relatively low to remove some degrees of correlation
Positive sample, so as to reduce the accumulated error that long-term follow is brought.In the position that degree of correlation maximum is determined after Multistage Detector
It puts and regards as target area in present frame, the method for reusing on-line study is updated object module, is increasing some just
Negative sample is to the training set of correlation filtering detector.A kind of re-detection response mechanism is finally established, when the maximum phase that detection obtains
During like spending small Mr. Yu's threshold value, detected again using not newer object module in target area is extended, search may go out
Existing target.Target location is determined after more than detection process, the overall process of target following in present frame is completed with this.
The present invention realizes that step is：
(1) build one by the cascade detector of Multistage Detector, and using detection result to target into line trace, often
The examination scope of level-one detector is descending, and similar to pyramidal inspection policies, reducing, the calculation amount per level-one is same
The accuracy of target is detected under the various environmental factors of Shi Tisheng；
(2) after a frame image sequence inputs cascade detector, as present frame, first with conspicuousness detector into
Row first order conspicuousness detects, if conspicuousness testing result value is more than threshold value, filters out several (for 5~10) roughly and waits
Target area is selected, background area in addition to this is removed；
(3) candidate target region for obtaining step (2) inputs second level grader set detector, is partitioned into fixed big
Small image block is selected 4~10 Harris corner features to be detected, using the principle of random binary tree, is calculated each
The corresponding posterior probability of angle point, judges whether posterior probability is more than the threshold value of setting, if more than the threshold value of setting, then will be greater than threshold
Input of the corresponding image block of posterior probability of value as third level detector；
(4) posterior probability obtained with step (3) is more than the image block of threshold value and negative sample positive in object module carries out phase
Guan Du is calculated, and first carries out being fourier transformed into frequency domain, convolution algorithm is converted to dot-product operation, significantly improves calculating speed, right
The degree of correlation result that positive sample calculates carries out descending sort, rejects the positive sample that the degree of correlation ranks behind, negative sample is calculated
Degree of correlation result carries out ascending sort, rejects the negative sample that the degree of correlation ranks behind；
(5) result of calculation of all positive negative sample degrees of correlation of traversal step (4) finds out the image block of maximum relation degree, i.e.,
For target position, and the positive and negative sample number deleted in the center surrounding sample of target position, supplement step (4),
And with this degree of correlation compared with the relevance threshold set, if less than the relevance threshold (relevance threshold 0.8) set,
Again it is detected, centered on the target location that the former frame of present frame detects, after 2 times of extensions are carried out to region of search,
Sliding window search is carried out to target in region after extension, redefines target；
It realizes so far and process is tracked to the target detection in current frame image, and the target to being likely to occur is lost
Situation problem takes the strategy detected again, completes the target following with automatic recovery ability based on Multistage Detector
Journey.
In the step (2), conspicuousness detector is：Image is first converted into Lab space, wherein L is brightness value, and a is
Again to the color value of bright pink from bottle green to grey, b is again to the color value of yellow, to all figures from sapphirine to grey
L, a, b channel of pixel are averaged as in, are set as Lm、am、bm, then the saliency value detector formula table of some pixel is calculated
It is shown as：
T=| | I(x,y)-Im||
S (x, y)=Uet
I in formula(x,y)For the corresponding brightness value of pixel that image coordinate is (x, y), ImBrightness for image entirety is averaged
Value, t be Euclidean distance that the pixel calculates as a result, U is fixed coefficient 0.5, S (x, y) is shown by what exponential function map
Work property numerical value；
The significant result acquired using above formula calculating, is set up the threshold value of one 0~255 and is judged, be partitioned into image
In significant region, be input to next stage detector as candidate target region, complete the detection process of conspicuousness detector.
In the step (3), the realization process of grader set detector is：
(1) it is carried first to choosing 4~10 Harris angle points in several candidate target regions determining in step (2)
It is that the average intensity change of generation is counted to the pixel progress sliding window translation (u, v) at (x, y) to take process：
Wherein I (x+u, y+v) and I (x, y) represents (x+u, y+v), the gray value at (x, y) respectively, and w (x, y) is Gauss
Weights of the weighting function at (x, y), E (u, v) are the changing value of pixel.
To above formula carry out Taylor expansion approximation be：
Wherein IxAnd IyFor the partial derivative in I (x, y) both direction.
Then covariance matrix therein is：
Two characteristic values of M are compared, angle point is regarded as if two characteristic values are all higher than the threshold value of setting, i.e., it is complete
Into the extraction process of an angle point；
(2) 4~10 Harris angle points of the procedure extraction are repeated, then count time that each angle point occurs in positive sample
Number scale is N, and the number occurred in negative sample is Q, then it is that the posterior probability of target area is to assert the angle point：
The posterior probability of gained is updated to random tree can be with express statistic posterior probability as a result, it is the every of random tree
The testing result of one layer of correspondence, one angle point, each node lower bifurcation go out two nodes, corresponding respectively whether to contain i-th jiao
Point is replaced with 1 and 0, after a complete random tree statistics, obtains a feature column vector by 0 and 1 combination, will
The posterior probability statistics of all situations is a normalized mapping table, when obtaining a feature vector knot with detection of classifier device
Fruit quickly searches the corresponding posterior probability of i-th of angle point by mapping；
Judge whether posterior probability is more than the threshold value of setting, if more than given threshold, then will be greater than the posterior probability of threshold value
Input of the corresponding image block as third level detector.The given threshold is with image texture correlation, and the present invention is by a large amount of
It is best that experiment, which is set as 0.01,.
In the step (4), the candidate target region that upper level is determined further is screened, these regions are made
For the input of correlation filtering detector, relatedness computation is carried out, selects the positive negative sample of initialization as initial target model, meter
The degree of correlation between the target area of input and positive negative sample is calculated, if the training set that positive negative sample is formed is：
WhereinRepresent positive sample,Represent negative sample, P*Represent the set that all positive negative samples are formed；
Then target area P1The degree of correlation with positive negative sample is respectively：
WhereinThe positive correlation degree of target area is represented,Represent the negatively correlated degree of target area；
Input target area is determined as the probability S of target1For：
In addition be ranked up to all with positive negative sample relatedness computation result, for the smaller positive sample of degree of correlation and
The larger negative sample of degree of correlation needs to delete.
The result of calculation of all positive negative sample degrees of correlation of traversal step (4), finds out maximum relation degree in the step (5)
Image block, as target position, and rejected just in the center surrounding sample of target position, supplement step (4)
The center of negative sample number, wherein positive sample sampling is：
Wherein n is the positive sample number rejected, and (x, y) be the target location center determined, the pixel of sliding window sampling step length λ=2,
The center of negative sample sampling is：
{∑(x±pλ,y±pλ)|p≥W+H,p∈Z}
Wherein m is the negative sample number rejected, and (x, y) be determining target location center, W and H be target area width and
It is long, the pixel of sliding window sampling step length λ=2；
And with maximum relation degree compared with the relevance threshold set, if less than the threshold value of the setting degree of correlation, weight is carried out
New detection, centered on the target location that the former frame of present frame detects, if detection zone spreading coefficient is pad：
Wherein SmaxFor maximum relation degree, ζ is relevance threshold, is 0.8.
Extended area sliding window is searched for using the object module of previous frame, the degree of correlation is calculated most using correlation filtering detector
Big region, then be newly defined as target.
The advantages of the present invention over the prior art are that：
(1) what the present invention was built be one by the cascade detector of Multistage Detector, and using detection result to target
Into line trace.The mechanism of this tracking can effectively overcome the limitation brought due to single detection method, improve various rings
The accuracy of target is detected under the factor of border.In addition detection range of the cascade detectors per level-one is descending, similar to pyramid
The inspection policies of shape can reduce calculation amount of every level-one in detection process, meet the reality in practical engineering application as far as possible
The requirement of when property.
(2) it is general to be updated to calculating posteriority in random tree in grader set detector using corner feature by the present invention
Rate, corner feature are the local features of image, row are remained unchanged to rotation, scaling, brightness change, to visual angle change.It is imitative
It penetrates variation, noise and also keeps a degree of stability.In addition the also unique good, abundant information of corner feature, diversity
The features such as, show preferable effect in terms of target is detected.And the multilayered structure of random tree can allow to select simultaneously it is multiple
Corner feature is detected judgement, will determine that result is configured to a feature vector, and posterior probability is obtained by mapping.Entire inspection
It is simple for structure efficiently to survey device, hence it is evident that shorten corner feature calculating and deterministic process, meet the accuracy of Practical Project and speed need
It asks.
(3) present invention utilizes the principle of correlation filtering in correlation filtering detector, it would be possible to which there are the candidate samples of target
Convolution algorithm, using the frequency-domain calculations degree of correlation is fourier transformed into, is converted to point multiplication operation by this with positive negative sample, is reduced related
The triviality of calculating further improves the calculating speed of entire detection process.In addition to candidate samples and each positive negative sample
Degree of correlation sort result rejects the relatively low sample of some degrees of correlation, so as to reduce the interference of accumulated error, prevents target following
There is drift phenomenon.
(4) present invention establishes a kind of re-detection mechanism, by the maximum relation degree of Multistage Detector output and given threshold ratio
Compared with if less than re-detection module is activated if the threshold value, centered on the target location of previous frame detection, in the region of search of extension
In to target carry out sliding window search, redefine target.This method overcomes to be caused during tracking since barrier blocks
Tracking failure problem, enhance the robustness of entire tracking system.
In short, the present invention, by the results show, this method has higher tracking accuracy than existing method, and possess weight
The automatic recovery ability of detection, while meet requirement of real-time in calculating speed, engineering technology is easily realized, therefore with practicability.
Description of the drawings
Fig. 1 is the target tracking algorism flow chart based on Multistage Detector；
Fig. 2 is the result figure of previous frame tracking box mark；
Fig. 3 is the schematic diagram detected to present frame conspicuousness；
Fig. 4 is the schematic diagram of Corner Detection successful match；
Fig. 5 is to carry out relatedness computation schematic diagram using positive sample；
Fig. 6 is the result figure of the final tracking box mark of present frame；
Fig. 7 once passes through Evaluation accuracy figure comparing result figure for the method for the present invention " Ours " with other algorithms.
Specific embodiment
The following describes the present invention in detail with reference to the accompanying drawings and embodiments.
As shown in Figure 1, present invention specific implementation step is as follows：
(1) conspicuousness detector is built
After first frame image is initialized, target position and area size are determined, and to the image block where target
Appropriate Gaussian noise and rotation transformation are superimposed, generates a series of target positive sample set, then the exterior domain random acquisition in target
The overall background image block such as several, as negative sample set.When after picture frame input after, first with conspicuousness detector pair
Marking area is detected in image.
The principle of conspicuousness detector is to calculate the average brightness value of each pixel and entire image all pixels point
Euclidean distance, it is marking area to find apart from the part of big Mr. Yu's threshold value after statistics.Concrete operations are first by input picture
XYZ space is transformed into, then be converted to Lab space by XYZ space by rgb space：
Wherein gamma () function representation is：
The formula for being transformed into XYZ space is expressed as：
The formula for turning Lab space by XYZ space again is expressed as：
After image is converted into Lab space, average value is acquired to L, a, b channel of all pixels respectively, is set as Lm、amWith
bm, then the saliency value formula for calculating some pixel is expressed as：
T=| | I(x,y)-Im||
S (x, y)=Uet
The S (x, y) calculated then regards as marking area if more than certain threshold value, therefore after the detection of conspicuousness detector
Image will appear multiple salient regions, exclude the interference of most of background, the search model of next stage detector will be inputed to
It encloses and further reduces.
(2) grader set detection module is built
In multiple salient regions in step (1) after the detection of conspicuousness detector, grader is further built
Gather detector.The thought of the grader set detector is calculated every mainly in conjunction with corner feature and random tree algorithm
The posterior probability of a salient region.Main processes of calculation is in initialized target region or positive sample region detection 4~10
Harris angle points, then same position goes out to detect whether that there are identical angle points in the current frame.Angle point calculating process be to
Pixel carries out sliding window translation (u, v) at (x, y), counts the average intensity change of generation：
Wherein I (x+u, y+v) and I (x, y) represents (x+u, y+v), the gray value at (x, y) respectively, and w (x, y) is Gauss
Weights of the weighting function at (x, y), E (u, v) are the changing value of pixel.
Consider Taylor expansion, above formula can approximation be reduced to：
Wherein IxAnd IyFor the partial derivative in I (x, y) both direction.
Then covariance matrix M is：
Determine whether Harris angle points, whether two characteristic values that need to only observe in M are larger, if more than a certain threshold value
Then regard as angle point.
4~10 angle points that the present invention selects all have fixed location tags for entire target image block, you can
It is detected whether at the same position of all positive negative samples there are this 4~10 corner features, if i-th of angle point is in all positive samples
There are n times altogether in middle detection, detects in all negative samples and occurs Q times altogether, then can assert and detect that i-th of angle point is target area
Posterior probability be：
4~10 all angle points are generalized to according to this thinking, then each angle point corresponds to a posterior probability.To working as
The salient region after the detection of conspicuousness detector in previous frame by step (1) carries out Corner Detection.Set up later one with
Machine tree is formed by several combination of nodes, each node is by condition judgment as a result, entering next two section of layer
One of them of point.The random tree that the present invention is set up shares 4~10 layers, corresponds to the testing result of each angle point respectively, often
A node lower bifurcation goes out two nodes, corresponds to target area respectively and contains i-th of angle point and without i-th of angle point, with 1 and 0 table
Show.After a complete random tree statistics, you can the feature column vector of one 10 × 1 is obtained, such as：
It is understood according to above-mentioned principle if 10 angle points, then all results of feature column vector have 210Kind, due to each
Node all corresponds to a posterior probability, can count the posterior probability of all situations for a normalized mapping table, when with
Detection of classifier device obtain a characteristic series vector result can by mapping quickly search its corresponding posterior probability, complete
Entire detection of classifier process, improves detection speed.
Posterior probability threshold value set as 0.5 according to many experiments measure, i.e., when the posterior probability of testing result is more than the threshold
Value, then may further determine that the marking area be target possibility it is very big, as the input of third level detector.
(3) correlation filtering detector is built
Using the multiple target areas filtered out in step (2) as the input of correlation filtering detector, degree of correlation meter is carried out
It calculates.Correlation filtering detector needs to build an object module, and the positive negative sample initialized in optional step (1) is as initial mesh
The training set of model is marked, calculates the degree of correlation between the target area of input and positive negative sample.The degree of correlation is convolution in time domain
Operation, but more simple point multiplication operation can be converted by being fourier transformed into frequency domain, by taking bivariate continuous signal as an example,
Specific form of calculation is as follows：
Dot product form can be written as by above formula in Fast Fourier Transform (FFT) to frequency domain：
H (τ, σ)=F (t, s) * G (t+ τ, s+ σ)
Using this principle, the relatedness computation between the target area of input and all positive negative samples is all transformed into frequency
It is carried out in domain, can largely reduce calculation amount, quickly obtain result.If the training set that positive negative sample is formed is：
Then target area P1The degree of correlation with positive negative sample is respectively：
So the probability that the input area is determined as target is：
In addition all result of calculations with the positive negative sample degree of correlation are ranked up, for the smaller positive sample of degree of correlation
Larger negative sample needs to delete with degree of correlation.
Because with the passage of tracking time, some original samples no longer adapt to the description of existing object module, if not
Object module sample set is updated, there is very big error with the degree of correlation that this sample set calculates, tracking can be caused to drift about.For
The influence of accumulated error is solved, needs to update sample set in each frame image, deletes the unconformable sample of some degrees of correlation.
(4) it determines target area and judges whether re-detection
By obtaining the corresponding probability value in multiple target areas in step (3), then the target area corresponding to probability value maximum
Target position in present frame is most likely to be, that is, completes detecting and tracking process in present frame.
If most probable value is less than certain threshold value of setting, illustrate in present frame target area by barrier block etc. because
Element and cause degree of correlation acute variation, then activate re-detection mechanism, if detection zone spreading coefficient be pad：
Wherein SmaxFor maximum relation degree, ζ is relevance threshold, value 0.8.
In re-detection mechanism, using the object module sample set of previous frame, the target location determined using previous frame is in
The heart, to extension target area sliding window search, whether detection target occurs again, judges that principle is identical with (3) to get to based on phase
The probability value of Guan Du, is compared with threshold value.
If the most probable value being calculated is more than threshold value again, illustrate to detect target area again, if being still below threshold
Value then continues re-detection.
(5) on-line study real-time update object module
Two parts are broadly divided into according to the update of on-line study object module.First part is the study based on time continuity
Update, because the target location in present frame is inevitable close with migrating the position into, according to what is deleted in step (3)
Positive sample number nearby extracts equivalent positive sample number as benefit using centered on the target location that previous frame determines in the current frame
It fills.
If the positive sample number deleted in (3) is n, previous frame target location center is (x, y), it is specified that the step-length of sliding window sampling
λ is 2 pixels, then sampling centre coordinate is：
Second part is considered based on spatiality, and only there are one may in same frame for synchronization in monotrack
Position, so according to the negative sample deleted in step (3), the target location determined in the current frame with previous frame is completely not
The extracted region equivalent negative sample of coincidence is as supplement.
If the negative sample number deleted in (3) is m, previous frame target's center and size are respectively (x, y), and W, H are, it is specified that sliding window
The step-length λ of sampling is 2 pixels, then sampling centre coordinate is：
{∑(x±pλ,y±pλ)|p≥W+H,p∈Z}
Pass through above two-part study, you can supplement is updated in real time to the positive negative sample in object module sample set,
So as to ensure that the reliability sum number purpose of positive negative sample is constant.
By above step, target position in present frame can be accurately positioned, utilize the cascade side of Multistage Detector
Formula can fast and effective accuracy of detection, have re-detection mechanism can target following failure in the case of detect mesh again
Mark, so as to restore target following, the process of on-line study can ensure positive and negative sample number while accumulated error is eliminated in addition
It is constant.Entire invention combines the thought of random tree and correlation filtering, largely reduces calculation amount, improves detection speed,
It is a kind of practicability algorithm haveing excellent performance.
As shown in Fig. 2, the people cycled in figure examines as a target in previous frame image by Multistage Detector
Its position is measured, and is marked and shown with tracking box.Start to adopt centered on the target location of the detection of previous frame in the current frame
Sample obtains a large amount of positive negative sample, the input as Multistage Detector.
As shown in figure 3, it is that the region of full figure conspicuousness is detected by conspicuousness detector first, later again with solid
Determine Threshold segmentation, it can be seen that there are 5 more apparent marking areas in this frame image, and respectively region area is different,
Using the candidate region of these targets as the input of next stage detector.
As shown in figure 4, carrying out Corner Detection to the candidate region of input by grader set detector, shown in figure every
One region can extract 4~10 angle points, count the probability that these angle points occur in positive negative sample, be updated to random tree
In, the possibility that the region is positive sample is calculated, with the threshold value comparison set up, its is corresponding if possibility is more than threshold value
Region is input to as candidate region in next stage detector.
As shown in figure 5, by correlation filtering detector, the candidate region of input and previous frame objective area in image are calculated
The degree of correlation, the corresponding region of the statistics maximum degree of correlation is determined as the target following result of present frame.
As shown in fig. 6, the detector by above three step detects successively, the mesh based on Multistage Detector is finally obtained
Tracking result is marked, and is marked and shown with tracking box.
As shown in fig. 7, be of the invention " Ours " with some other current mainstream algorithm (" MEEM ", " SAM ", " KCF ",
" Struct ", " TLD ") comparing result, in figure title " Precision plots of OPE " expression once pass through assessment
Accuracy, abscissa " Location error threshold " represent zone errors threshold value, and ordinate " Precision " represents
Accuracy.
Above example is provided just for the sake of the description purpose of the present invention, and is not intended to limit the scope of the present invention.This
The range of invention is defined by the following claims.It the various equivalent replacements that do not depart from spirit and principles of the present invention and make and repaiies
Change, should all cover within the scope of the present invention.
Claims (9)
1. a kind of method for tracking target with automatic recovery ability based on Multistage Detector, it is characterised in that：Step is as follows：
(1) one is built by the cascade detector of Multistage Detector, and utilizes the result detected to target into line trace, per level-one
The examination scope of detector is descending, similar to pyramidal inspection policies, is carried simultaneously reducing the calculation amount per level-one
Rise the accuracy that target is detected under various environmental factors；
(2) after a frame image sequence inputs cascade detector, as present frame, the is carried out first with conspicuousness detector
Level-one conspicuousness detects, if conspicuousness testing result value is more than threshold value, several candidate target regions is filtered out roughly, except this
Except background area be removed；
(3) candidate target region for obtaining step (2) inputs second level grader set detector, is partitioned into fixed size figure
As block, 4~10 Harris corner features is selected to be detected, using the principle of random binary tree, each angle point is calculated
Corresponding posterior probability, finally judges whether total posterior probability is more than the threshold value of setting, then will be big if more than the threshold value of setting
In input of the corresponding image block of the posterior probability of threshold value as third level detector；
(4) positive negative sample in the image block and object module of threshold value is more than with the posterior probability that step (3) obtains and carries out the degree of correlation
It calculates, first carries out being fourier transformed into frequency domain, convolution algorithm is converted into dot-product operation, significantly improves calculating speed, to positive sample
The degree of correlation result of this calculating carries out descending sort, rejects the positive sample that the degree of correlation ranks behind, the correlation calculated negative sample
It spends result and carries out ascending sort, reject the negative sample that the degree of correlation ranks behind；
(5) result of calculation of all positive negative sample degrees of correlation of traversal step (4) finds out the image block of maximum relation degree, as mesh
Mark position, and the positive and negative sample number deleted in the center surrounding sample of target position, supplement step (4), and with
If this degree of correlation less than the relevance threshold of setting, is detected, compared with the relevance threshold ζ set with current again
Centered on the target location of the former frame detection of frame, after 2 times of extensions are carried out to region of search, to target in region after expansion
Sliding window search is carried out, redefines target；
The process that tracked to the target detection in current frame image, and the target loss situation to being likely to occur are realized so far
Problem takes the strategy detected again, completes the object tracking process with automatic recovery ability based on Multistage Detector.
2. the method for tracking target with automatic recovery ability according to claim 1 based on Multistage Detector, feature
It is：In the step (2), conspicuousness detector is：Image is first converted into Lab space, wherein L is brightness value, and a is from depth
Green arrives the color value of bright pink to grey again, and b is again to the color value of yellow, in all images from sapphirine to grey
L, a, b channel of pixel are averaged, and are set as Lm、am、bm, then the saliency value detector formula for calculating some pixel is expressed as：
T=| | I(x,y)-Im||
S (x, y)=Uet
I in formula(x,y)For the corresponding brightness value of pixel that image coordinate is (x, y), ImFor the average brightness of image entirety, t
For Euclidean distance for pixel calculating as a result, U is fixed coefficient, S (x, y) is the conspicuousness number mapped by exponential function
Value；
The significant result acquired using above formula calculating, is set up the threshold value of one 0~255 and is judged, is partitioned into image and is shown
The region of work is input to next stage detector as candidate target region, completes the detection process of conspicuousness detector.
3. the method for tracking target with automatic recovery ability according to claim 1 based on Multistage Detector, feature
It is：Several in the step (2) are 5~10.
4. the method for tracking target with automatic recovery ability according to claim 1 based on Multistage Detector, feature
It is：The fixed coefficient U is 0.5.
5. the method for tracking target with automatic recovery ability according to claim 1 based on Multistage Detector, feature
It is：In the step (3), the realization process of grader set detector is：
(1) it was extracted first to choosing 4~10 Harris angle points in several candidate target regions determining in step (2)
Journey is that the average intensity change of generation is counted to the pixel progress sliding window translation (u, v) at (x, y)：
Wherein I (x+u, y+v) and I (x, y) represents (x+u, y+v), the gray value at (x, y) respectively, and w (x, y) is Gauss weighting
Weights of the function at (x, y), E (u, v) are the changing value of pixel；
To above formula carry out Taylor expansion approximation be：
Wherein IxAnd IyFor the partial derivative in I (x, y) both direction,
Then covariance matrix therein is：
Two characteristic values of M are compared, angle point is regarded as if two characteristic values are all higher than the threshold value of setting, that is, completes one
The extraction process of a angle point；
(2) 4~10 Harris angle points of the procedure extraction are repeated, then count the secondary number scale that each angle point occurs in positive sample
For N, the number occurred in negative sample is Q, then it is that the posterior probability of target area is to assert the angle point：
The posterior probability of gained is updated to random tree express statistic posterior probability as a result, it is each layer of correspondence of random tree
Whether the testing result of one angle point, each node lower bifurcation go out two nodes, corresponding containing i-th of angle point respectively, with 1 and 0
Instead of after a complete random tree statistics, a feature column vector by 0 and 1 combination being obtained, by all situations
Posterior probability statistics is a normalized mapping table, when obtaining a feature vector result with detection of classifier device and pass through to reflect
It penetrates and quickly searches the corresponding posterior probability of i-th of angle point；
Judge whether posterior probability is more than the threshold value of setting, if more than given threshold, then the posterior probability that will be greater than threshold value corresponds to
Input of the image block as third level detector.
6. the method for tracking target with automatic recovery ability according to claim 5 based on Multistage Detector, feature
It is：The given threshold is 0.01.
7. the method for tracking target with automatic recovery ability according to claim 1 based on Multistage Detector, feature
It is：In the step (4), the candidate target region that upper level is determined further is screened, using these regions as
The input of correlation filtering detector carries out relatedness computation, and the positive negative sample of initialization is selected to be calculated as initial target model
The degree of correlation between the target area of input and positive negative sample, if the training set that positive negative sample is formed is：
WhereinRepresent positive sample,Represent negative sample, P*Represent the set that all positive negative samples are formed；
Then target area P1The degree of correlation with positive negative sample is respectively：
WhereinThe positive correlation degree of target area is represented,Represent the negatively correlated degree of target area；
Input target area is determined as the probability S of target1For：
In addition it is ranked up to all with positive negative sample relatedness computation result：The degree of correlation result calculated positive sample drops
Sequence sorts, and rejects the positive sample that the degree of correlation ranks behind, and carries out ascending sort to the degree of correlation result that negative sample calculates, rejects phase
The negative sample that Guan Du ranks behind.
8. the method for tracking target with automatic recovery ability according to claim 1 based on Multistage Detector, feature
It is：The result of calculation of all positive negative sample degrees of correlation of traversal step (4) in the step (5) finds out the figure of maximum relation degree
As block, as target position, and rejected in the center surrounding sample of target position, supplement step (4) positive and negative
The center of sample number, wherein positive sample sampling is：
Wherein n is the positive sample number rejected, and (x, y) be determining target location center, the pixel of sliding window sampling step length λ=2, negative sample
The center of this sampling is：
{∑(x±pλ,y±pλ)|p≥W+H,p∈Z}
Wherein m is the negative sample number rejected, and (x, y) is determining target location center, and W and H are the wide with long of target area, sliding
The pixel of window sampling step length λ=2；
And with maximum relation degree compared with the relevance threshold set, if less than the threshold value of the setting degree of correlation, examined again
It surveys, centered on the target location that the former frame of present frame detects, if detection zone spreading coefficient is pad：
Wherein SmaxFor maximum relation degree, ζ is relevance threshold；
Extended area sliding window is searched for using the object module of previous frame, degree of correlation maximum is calculated using correlation filtering detector
Region is then newly defined as target.
9. the method for tracking target with automatic recovery ability based on Multistage Detector according to claim 1 or 8, special
Sign is：In the step (5), relevance threshold ζ is 0.8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711422462.2A CN108154159B (en) | 2017-12-25 | 2017-12-25 | A kind of method for tracking target with automatic recovery ability based on Multistage Detector |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711422462.2A CN108154159B (en) | 2017-12-25 | 2017-12-25 | A kind of method for tracking target with automatic recovery ability based on Multistage Detector |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108154159A true CN108154159A (en) | 2018-06-12 |
| CN108154159B CN108154159B (en) | 2018-12-18 |
Family
ID=62462003
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711422462.2A Active CN108154159B (en) | 2017-12-25 | 2017-12-25 | A kind of method for tracking target with automatic recovery ability based on Multistage Detector |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108154159B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109325520A (en) * | 2018-08-24 | 2019-02-12 | 北京航空航天大学 | A kind of inspection method of Oil spills, apparatus and system |
| CN109558505A (en) * | 2018-11-21 | 2019-04-02 | 百度在线网络技术（北京）有限公司 | Visual search method, apparatus, computer equipment and storage medium |
| CN110046659A (en) * | 2019-04-02 | 2019-07-23 | 河北科技大学 | A kind of long-time monotrack method based on TLD |
| CN111327833A (en) * | 2020-03-31 | 2020-06-23 | 厦门美图之家科技有限公司 | Auxiliary shooting method and device, electronic equipment and readable storage medium |
| CN111627046A (en) * | 2020-05-15 | 2020-09-04 | 北京百度网讯科技有限公司 | Target part tracking method and device, electronic equipment and readable storage medium |
| CN112102356A (en) * | 2019-06-18 | 2020-12-18 | 北京七鑫易维科技有限公司 | Target tracking method and device, terminal equipment and storage medium |
| CN112890572A (en) * | 2021-02-07 | 2021-06-04 | 广州一盒科技有限公司 | Intelligent control system and method for cooking food materials |
| CN116778532A (en) * | 2023-08-24 | 2023-09-19 | 汶上义桥煤矿有限责任公司 | Underground coal mine personnel target tracking method |
| CN117635637A (en) * | 2023-11-28 | 2024-03-01 | 北京航空航天大学 | Autonomous conceived intelligent target dynamic detection system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7848548B1 (en) * | 2007-06-11 | 2010-12-07 | Videomining Corporation | Method and system for robust demographic classification using pose independent model from sequence of face images |
| CN104301585A (en) * | 2014-09-24 | 2015-01-21 | 南京邮电大学 | Method for detecting specific kind objective in movement scene in real time |
| CN107358623A (en) * | 2017-07-12 | 2017-11-17 | 武汉大学 | A kind of correlation filtering track algorithm based on conspicuousness detection and robustness size estimation |
-
2017
- 2017-12-25 CN CN201711422462.2A patent/CN108154159B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7848548B1 (en) * | 2007-06-11 | 2010-12-07 | Videomining Corporation | Method and system for robust demographic classification using pose independent model from sequence of face images |
| CN104301585A (en) * | 2014-09-24 | 2015-01-21 | 南京邮电大学 | Method for detecting specific kind objective in movement scene in real time |
| CN107358623A (en) * | 2017-07-12 | 2017-11-17 | 武汉大学 | A kind of correlation filtering track algorithm based on conspicuousness detection and robustness size estimation |
Non-Patent Citations (4)
| Title |
|---|
| BHORAM LEE 等: "Self-Supervised Online Learning of Appearance for 3D Tracking", 《INTELLIGENT ROBOTS AND SYSTEMS》 * |
| PETER YUEN 等: "A cortex like neuromorphic target recognition & tracking in cluttered background", 《IMAGING FOR CRIME DETECTION AND PREVENTION》 * |
| YACHUN FENG 等: "Multi-expert tracking algorithm based on improved compressive tracker", 《INTERNATIONAL CONFERENCE ON GRAPHIC AND IMAGE PROCESSING》 * |
| 王海地: "基于显著性检测和分类器训练的航拍图像车辆检测", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109325520A (en) * | 2018-08-24 | 2019-02-12 | 北京航空航天大学 | A kind of inspection method of Oil spills, apparatus and system |
| US11348254B2 (en) | 2018-11-21 | 2022-05-31 | Baidu Online Network Technology (Beijing) Co., Ltd. | Visual search method, computer device, and storage medium |
| CN109558505A (en) * | 2018-11-21 | 2019-04-02 | 百度在线网络技术（北京）有限公司 | Visual search method, apparatus, computer equipment and storage medium |
| CN110046659A (en) * | 2019-04-02 | 2019-07-23 | 河北科技大学 | A kind of long-time monotrack method based on TLD |
| CN110046659B (en) * | 2019-04-02 | 2023-04-07 | 河北科技大学 | TLD-based long-time single-target tracking method |
| CN112102356A (en) * | 2019-06-18 | 2020-12-18 | 北京七鑫易维科技有限公司 | Target tracking method and device, terminal equipment and storage medium |
| CN111327833A (en) * | 2020-03-31 | 2020-06-23 | 厦门美图之家科技有限公司 | Auxiliary shooting method and device, electronic equipment and readable storage medium |
| CN111327833B (en) * | 2020-03-31 | 2021-06-01 | 厦门美图之家科技有限公司 | Auxiliary shooting method and device, electronic equipment and readable storage medium |
| CN111627046A (en) * | 2020-05-15 | 2020-09-04 | 北京百度网讯科技有限公司 | Target part tracking method and device, electronic equipment and readable storage medium |
| CN112890572A (en) * | 2021-02-07 | 2021-06-04 | 广州一盒科技有限公司 | Intelligent control system and method for cooking food materials |
| CN116778532A (en) * | 2023-08-24 | 2023-09-19 | 汶上义桥煤矿有限责任公司 | Underground coal mine personnel target tracking method |
| CN116778532B (en) * | 2023-08-24 | 2023-11-07 | 汶上义桥煤矿有限责任公司 | Underground coal mine personnel target tracking method |
| CN117635637A (en) * | 2023-11-28 | 2024-03-01 | 北京航空航天大学 | Autonomous conceived intelligent target dynamic detection system |
| CN117635637B (en) * | 2023-11-28 | 2024-06-11 | 北京航空航天大学 | Autonomous conceived intelligent target dynamic detection system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108154159B (en) | 2018-12-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108154159B (en) | A kind of method for tracking target with automatic recovery ability based on Multistage Detector | |
| Li et al. | Automatic pavement crack detection by multi-scale image fusion | |
| Yang et al. | Real-time face detection based on YOLO | |
| CN106127204B (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN106096577B (en) | A kind of target tracking method in camera distribution map | |
| CN102542289B (en) | Pedestrian volume statistical method based on plurality of Gaussian counting models | |
| Miao et al. | An object-based method for road network extraction in VHR satellite images | |
| CN110135296A (en) | Airfield runway FOD detection method based on convolutional neural networks | |
| Liu et al. | A contrario comparison of local descriptors for change detection in very high spatial resolution satellite images of urban areas | |
| CN103886325B (en) | Cyclic matrix video tracking method with partition | |
| CN107145862A (en) | A kind of multiple features matching multi-object tracking method based on Hough forest | |
| CN108198201A (en) | A kind of multi-object tracking method, terminal device and storage medium | |
| CN107507170A (en) | A kind of airfield runway crack detection method based on multi-scale image information fusion | |
| CN110298297A (en) | Flame identification method and device | |
| CN105160355B (en) | A kind of method for detecting change of remote sensing image based on region correlation and vision word | |
| CN104077605A (en) | Pedestrian search and recognition method based on color topological structure | |
| CN108648211A (en) | A kind of small target detecting method, device, equipment and medium based on deep learning | |
| CN109886356A (en) | A kind of target tracking method based on three branch's neural networks | |
| CN108460790A (en) | A kind of visual tracking method based on consistency fallout predictor model | |
| CN110991397B (en) | Travel direction determining method and related equipment | |
| CN110097091A (en) | It is trained be distributed with inference data it is inconsistent under the conditions of image fine granularity recognition methods | |
| CN106778570A (en) | A kind of pedestrian's real-time detection and tracking | |
| CN113111727A (en) | Method for detecting rotating target in remote sensing scene based on feature alignment | |
| CN106709938A (en) | Multi-target tracking method based on improved TLD (tracking-learning-detected) | |
| Wang et al. | Low-altitude infrared small target detection based on fully convolutional regression network and graph matching |
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 |