CN104156734B - A kind of complete autonomous on-line study method based on random fern grader - Google Patents
A kind of complete autonomous on-line study method based on random fern grader Download PDFInfo
- Publication number
- CN104156734B CN104156734B CN201410407669.2A CN201410407669A CN104156734B CN 104156734 B CN104156734 B CN 104156734B CN 201410407669 A CN201410407669 A CN 201410407669A CN 104156734 B CN104156734 B CN 104156734B
- Authority
- CN
- China
- Prior art keywords
- sample
- positive
- random fern
- grader
- line study
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of complete autonomous on-line study method based on random fern grader, the method only needs the frame in the video frame the grader on-line study for the target class is carried out by selecting a target.Step is：Target first to frame choosing obtains initial positive sample collection using affine transformation, and a small amount of initial random fern grader of negative sample collection training is extracted in the nontarget area of video；Secondly, target detection is carried out in the video frame using the grader.During detection, on-line study new samples, and automatic decision sample class are collected using nearest neighbor classifier；Finally, new samples are used for the on-line training of random fern grader, random fern posterior probability is updated, the precision of grader target detection is gradually stepped up, the complete autonomous on-line study of object detection system is realized.
Description
Technical field
The invention belongs to area of pattern recognition, and in particular to a kind of complete autonomous on-line study side based on random fern grader
Method.
Background technology
On-line study belongs to the research category of incremental learning, and grader only learns one to each sample in this class method
It is secondary, rather than the study for repeating, do not need substantial amounts of memory space to store training in such on-line learning algorithm running
Sample, grader often obtains a sample, i.e., carry out on-line study to it, and grader is made in use by on-line study
Still classifying quality can further be improved according to new samples self-renewing and improvement.
The on-line learning algorithm of early stage has a Winnow algorithms, unified linear prediction algorithm etc., scholar Oza in 2001 by these
Algorithm is combined with boosting algorithms, it is proposed that (algorithm is quoted from " Online bagging for online boosting algorithms
And boosting " N.Oza and S.Russell, In Proc.Artificial Intelligence and
Statistics, 105-112,2001), in the method for Oza, each feature one Weak Classifier of correspondence, and strong classifier is
The weighted sum of a number of Weak Classifier, wherein Weak Classifier are all select from weak classifier set.It is online to learn
During habit, each Weak Classifier in each training sample renewal weak classifier set one by one, including adjust dividing for positive negative sample
The weight of class threshold value and the grader, the Weak Classifier weight more and more higher for having made, and poor Weak Classifier weight is more next
It is lower, so that one sample of on-line study can just pick out a strong classification of present weight highest Weak Classifier addition every time
Making the grader for finally training in device has stronger classification capacity.
But, each Weak Classifier will be carried out online to new samples in the weak classifier set of online boosting algorithms
Study, when Weak Classifier number is more, on-line study speed will necessarily be slack-off.Grabner enters to online boosting algorithms
Go improvement, it is equally carried out feature selecting as well as Adaboost algorithm, and this feature selecting and to classification
More new capital of device is to carry out online, and (algorithm is quoted from " On-line boosting and for referred to as online Adaboost
Vision " H.Grabner and H.Bischof, In Proc.CVPR, (1):260-267,2006).But it is online
Adaboost replaces general Weak Classifier to synthesize strong classifier with feature selecting operator, and feature selecting operator number and feature are selected
It is all fixed to select the corresponding Weak Classifier number of operator, and corresponding on-line study grader structure relatively ossifys.When its point of discovery
When class ability cannot meet the requirement of detection performance, even if lasting on-line study is gone down cannot also improve accuracy of detection.
Ozuysal does not use Weak Classifier composition strong classifier and carries out sample classification, but from sample feature set with
Machine extracts multiple features and constitutes a random fern, and training sample Posterior probability distribution is counted by random fern, then by multiple random
The Posterior probability distribution of fern carries out sample classification, and as (algorithm is quoted from " Fast keypoint for random fern classifier algorithm
Recognition using random ferns " In Pattern Analysis Machine Intelligence, IEEE
Transaction on 32(3),448-461,2010)。
The content of the invention
The technical problem to be solved in the present invention is：A kind of complete autonomous on-line study side based on random fern grader is provided
Method, for the autonomous learning of grader improving classification performance.
The present invention is for the solution technical scheme taken of above-mentioned technical problem：It is a kind of based on random fern grader it is complete from
Main on-line study method, it is characterised in that：It comprises the following steps：
1) sample set of initial training grader is prepared：
For frame of video to be detected, frame selects a Target Photo in the video frame, the Target Photo is carried out affine
The picture that conversion is obtained is used as positive sample；Not contain the background image region of target as negative sample；So random acquisition
A number of positive sample and negative sample as initial training grader sample set；Positive and negative samples are size identical image
Block；
2) random fern grader initial training：
Initial training is carried out to random fern grader using the sample set of ready initial training grader, is counted positive and negative
Posterior probability distribution of the sample in each random fern；
3) the good random fern grader of initial training is traveled through into frame of video to be detected as current goal detector is carried out
Target detection, obtains object module, and calculate the confidence level of each object module；
4) positive and negative sample form collection is built：
Following three kinds of samples are added to positive sample template set M as positive sample template+, remaining is added to negative sample template
Collection M-：
A, step 1) in the positive sample that obtains；
B, to step 3) in the confidence level that obtains exceed the object module of confidence level preset value, using optical flow method to where it
Frame of video is tracked and obtains tracking module, if tracking module has overlapping region, and coincidence factor more than default with the object module
Coincidence factor, then it is assumed that the tracking module is real goal, M is added to as positive sample template+In；
C, to step 3) in the confidence level that obtains exceed the object module of confidence level preset value, using optical flow method to where it
Frame of video is tracked and obtains tracking module, if tracking module has overlapping region, and coincidence factor not less than pre- with the object module
If coincidence factor, then by conservative similarity ScJudge that can the tracking module add positive sample template set：
Wherein：
If ScMore than default conservative similarity threshold, then the tracking module is used as positive sample template addition M+,For
The similarity of sample to be sorted and the first half template of current positive sample template set, S+、S-Sample respectively to be sorted with just,
The similarity of negative sample template set,It is two similarities of picture frame, p+, p-Respectively positive sample and negative sample, p is
Sample to be sorted, sample to be sorted is tracking module in this step；
A positive sample template is often added, then taking in same frame of video four image blocks of formed objects around it and judging is
No is negative sample, if adding negative sample template set M as negative sample template-；
5) nearest neighbor classifier is used, the positive negative sample of on-line study is obtained：
The setting of nearest neighbor classifier is as follows：For each sample p to be sorted, it is calculated respectively with positive and negative sample form collection
Similarity S+(p,M+) and S-(p,M-)：
Similarity S can be obtained accordinglyr：
If similarity SrMore than threshold θNN, then judge that the sample to be sorted is real goal, as the positive sample of on-line study
This；It is otherwise false-alarm, as the negative sample of on-line study；
Sample to be sorted is step 3 in this step) object module and step 4 that obtain) the positive and negative sample form collection that obtains；
(6) on-line training of random fern grader：
Using step 5) obtain on-line study positive negative sample, on-line study is carried out to random fern grader, gradually carry
Its nicety of grading high；
Target detection is carried out using the random fern grader of on-line study as the detecting system of sustainable renewal.
By such scheme, described step 2) specific method it is as follows：
2.1) random fern is constructed：
To taking s to characteristic point as one group of random fern at random in the single sample in the sample set of initial training grader,
The position that each sample takes characteristic point is identical, and each pair characteristic point carries out the comparing of pixel value, previous feature in each pair characteristic point
Point pixel value then takes greatly characteristic value for 1, on the contrary then take s characteristic value that characteristic value obtains for 0, s more afterwards to characteristic point according to
Random order constitutes the binary number of s, as this group random fern numerical value of random fern, in the random fern of each sample
The sequence consensus of characteristic value；
2.2) posterior probability of the random fern numerical value in positive negative sample class is calculated：
In random fern, some is obtained for positive sample, and other are obtained for negative sample；The value kind of random fern numerical value
Class has 2sIt is individual；
The positive sample number of the value of every kind of random fern numerical value is counted, so as to obtain random fern numerical value in positive sample class C1On
Posterior probability distribution P (Fl|C1)；Random fern numerical value is similarly obtained in negative sample class C0On Posterior probability distribution P (Fl|C0)；
Combine all random ferns to classify the sample set of initial training grader, as random fern grader；
Described step 3) target detection is carried out in every frame video image using above-mentioned random fern grader：
Traversal every frame video image to be detected, extracts the image block of formed objects as to be measured in every frame video image
Sample, size and the step 1 of sample to be tested) in positive sample it is equal in magnitude, calculate the random fern numerical value of each sample to be tested, from
And corresponding posterior probability is obtained, finally by its classification of random fern classifier calculated；
It is the image block of positive sample for classification, then is detected as target.
By such scheme, described step 4) often add a positive sample template, then take in same frame of video four around it
When the image block of individual formed objects determines whether negative sample, Gaussian Background modeling is introduced, if foreground pixel is less than in image block
Foreground pixel threshold value, then judge that it is negative sample.
By such scheme, described step 4) also include that template set cuts down mechanism：Sample to be sorted and positive and negative template set
Maximum of the similarity equal to similarity between single positive and negative sample form in sample to be sorted and positive and negative template set；Real-time statistics
Each positive and negative sample form obtains the number of times of the maximum, if the number of times of the maximum of certain positive and negative sample form acquisition is less than most
Big value number of times preset value, then remove corresponding positive sample template or negative sample template.
By such scheme, described step 6) on-line study of random fern grader is by updating Posterior probability distribution reality
It is existing.
By such scheme, described step 6) specific method is as follows：
6.1) using step 5) the positive negative sample that obtains is used as on-line study sample；If an on-line study sample is (fnew,
ck), wherein fnewIt is the binary number of random fern s, ckIt is sample class, calculates the random fern numerical value of the on-line study sample；
6.2) to step 2.1) classification is c in sample setkTotal sample number add 1, classification is ckWith the on-line study sample
Random fern numerical value identical sample number add 1；The sample number of other random fern numerical value is constant；
6.3) according to the sample number after renewal, Posterior probability distribution of the random fern numerical value in the sample class is recalculated；
6.4) an on-line study sample is often increased newly, just repeatedly 6.1) to 6.3) being updated one to Posterior probability distribution
It is secondary.
Beneficial effects of the present invention are：
1st, only needing frame in the video frame carries out the grader on-line study for the target class by selecting a target, i.e.,：
Target first to frame choosing obtains initial positive sample collection using affine transformation, and a small amount of bearing is extracted in the nontarget area of video
Sample set trains initial random fern grader；Secondly, target detection is carried out in the video frame using the grader；The process of detection
In, on-line study new samples, and automatic decision sample class are collected using nearest neighbor classifier；Finally, by the new sample of on-line study
This is used for the on-line training of random fern grader, updates random fern posterior probability, gradually steps up random fern grader target detection
Precision, realize the complete autonomous on-line study of object detection system.
2nd, this patent introduces template set abatement mechanism, and in can avoiding template set, positive and negative sample form is more to be likely to result in
The shortcoming that system running speed declines.
Brief description of the drawings
Fig. 1 is the flow chart of the inventive method；
Fig. 2 is random fern grader structure chart in one embodiment of the invention；
Fig. 3 is the comparison diagram of detection performance before and after the random fern grader of one embodiment of the invention on-line study, wherein Fig. 3
A () -3 (d) is the testing result before on-line study, Fig. 3 (i) -3 (l) is the testing result after on-line study；
Fig. 4 is the grader autonomous learning procedure chart under night illumination condition；
Fig. 5 is the grader autonomous learning procedure chart of pedestrian detection；
Fig. 6 compares figure for one embodiment of the invention with the ROC curve of other classical on-line study processes.
Specific embodiment
With reference to instantiation and accompanying drawing, the present invention will be further described.
The invention discloses the nearest neighbor classifier during the complete autonomous on-line study studied based on object detection system
Training method, the grader that the method only needs frame in the video frame and carries out being directed to the target class by selecting a target is learned online
Practise.Step is：Target first to frame choosing obtains initial positive sample collection using affine transformation, is carried in the nontarget area of video
Take a small amount of initial random fern grader of negative sample collection training；Secondly, target detection is carried out in the video frame using the grader.
During detection, on-line study new samples, and automatic decision sample class are collected using nearest neighbor classifier；Finally, will be new
Sample is used for the on-line training of random fern grader, updates random fern posterior probability, gradually steps up the essence of grader target detection
Degree, realizes the complete autonomous on-line study of object detection system.
The present invention provides a kind of complete autonomous on-line study method based on random fern grader as shown in figure 1, including as follows
Step：
1) sample set of initial training grader is prepared：
For frame of video to be detected, a target is selected in the first frame center of video image, the Target Photo is entered
The picture that row affine transformation is obtained is used as positive sample；Not contain the background image region of target as negative sample；It is so random
The a number of positive sample of acquisition and negative sample as initial training grader sample set.
Sample in the sample set of the initial training grader is exactly in the present embodiment the image block of formed objects, one
As size be 15 × 15 (pixels), if containing the sample if target to be detected being positive sample in image block, it is no then to bear sample
This.
2) random fern grader initial training：
Initial training is carried out to random fern grader using the sample set of ready initial training grader, is counted positive and negative
Posterior probability distribution of the sample in each random fern, as shown in Figure 2.
Specific method is as follows：
2.1) random fern is constructed：
S is taken to characteristic point as one group of random fern (the present embodiment selects 5 pairs) to random in the single sample in sample set, often
The position that individual sample takes characteristic point is identical, and each pair characteristic point carries out the comparing of pixel value, previous characteristic point in each pair characteristic point
Pixel value then takes greatly characteristic value for 1, on the contrary then take s characteristic value that characteristic value obtains for 0, s more afterwards to characteristic point according to
The order of machine constitutes the binary number of s, and as this group random fern numerical value of random fern is special in the random fern of each sample
The sequence consensus of value indicative；
2.2) posterior probability of the random fern numerical value in positive negative sample class is calculated：
In random fern, some is obtained for positive sample, and other are obtained for negative sample；The random fern F of each samplel
Comprising feature can be united to form a decimal number, because the decimal number is obtained by S binary code, therefore
The value species of random fern numerical value has 2sIt is individual, that is, have 2sPlant (to be 2 in the present embodiment5Plant possible)；
The positive sample number of the value of every kind of random fern numerical value is counted, so as to obtain random fern numerical value in positive sample class C1On
Posterior probability distribution P (Fl|C1)；Random fern numerical value is similarly obtained in negative sample class C0On Posterior probability distribution P (Fl|C0)；
Combine all random ferns to classify the sample set of initial training grader, as random fern grader.
3) the good random fern grader of initial training is traveled through into frame of video to be detected as current goal detector is carried out
Target detection, obtains object module, and calculates the confidence level of each object module, specially：Traversal frame of video to be detected,
The image block of formed objects is extracted in frame of video as sample to be tested, size and the step 1 of sample to be tested) in positive sample size
It is equal, the random fern numerical value of each sample to be tested is calculated, so as to obtain corresponding posterior probability, finally by random fern grader meter
Calculate its classification；
It is the image block of positive sample for classification, then is detected as target, as object module.
4) positive and negative sample form collection is built：
Following three kinds of samples are added to positive sample template set M as positive sample template+, remaining is added to negative sample template
Collection M-：
A, step 1) in the positive sample that obtains；
B, to step 3) in the confidence level that obtains exceed the object module of confidence level preset value (desirable 0.6), using light stream
Method is tracked to frame of video where it and obtains tracking module, if tracking module has overlapping region with the object module, and overlaps
Rate exceedes default coincidence factor (default coincidence factor generally takes 60%), then it is assumed that the tracking module is real goal, used as positive sample
Template is added to M+In；
C, to step 3) in the confidence level that obtains exceed the object module of confidence level preset value (desirable 0.6), using light stream
Method is tracked to frame of video where it and obtains tracking module, if tracking module has overlapping region with the object module, and overlaps
Rate not less than default coincidence factor, then by conservative similarity ScJudge that can the tracking module add positive sample template set：
Wherein：
If ScMore than default conservative similarity threshold (desirable 0.6), then the tracking module adds as positive sample template
Enter M+,It is sample to be sorted and the similarity of the first half template of current positive sample template set, S+、S-Respectively treat point
The similarity of class sample and positive and negative samples template set,It is two similarities of picture frame, p+, p-Respectively positive sample
And negative sample, p is sample to be sorted, and sample to be sorted is tracking module in this step；
A positive sample template is often added, then taking in same frame of video four image blocks of formed objects around it and judging is
No is negative sample, if adding negative sample template set M as negative sample template-.When judging, Gaussian Background modeling is introduced, if
Foreground pixel then judges that it is negative sample less than foreground pixel threshold value (desirable to be less than 30%) in image block.
Step 4) also include that template set cuts down mechanism：Sample to be sorted is equal to sample to be sorted with the similarity of positive and negative template set
In this and positive and negative template set between single positive and negative sample form similarity maximum；Real-time statistics each positive and negative sample forms is obtained
The number of times of the maximum is obtained, if the number of times of the maximum of certain positive and negative sample form acquisition is less than maximum number of times preset value,
Remove corresponding positive sample template or negative sample template.
5) nearest neighbor classifier is used, the positive negative sample of on-line study is obtained：
The setting of nearest neighbor classifier is as follows：For each sample p to be sorted, it is calculated respectively with positive and negative sample form collection
Similarity S+(p,M+) and S-(p,M-)：
Similarity S can be obtained accordinglyr：
If similarity SrMore than threshold θNN, then judge that the sample to be sorted is real goal, as the positive sample of on-line study
This；It is otherwise false-alarm, as the negative sample of on-line study；
Sample to be sorted is step 3 in this step) object module and step 4 that obtain) the positive and negative sample form collection that obtains.
(6) on-line training of random fern grader：
Using step 5) obtain on-line study positive negative sample, on-line study is carried out to random fern grader, gradually carry
Its nicety of grading high；Target detection is carried out using the random fern grader of on-line study as the detecting system of sustainable renewal.
The on-line study of random fern grader realizes that specific method is as follows by updating Posterior probability distribution：
6.1) using step 5) the positive negative sample that obtains is used as on-line study sample；If an on-line study sample is (fnew,
ck), wherein fnewIt is (the f in the present embodiment of the binary number of random fern snewBe 00101, i.e. decimal number 5), ckIt is sample class
Not, the random fern numerical value of the on-line study sample is calculated；
6.2) as shown in Fig. 2 to step 2.1) classification is c in sample setkTotal sample number add 1, classification is ckWith this
The random fern numerical value identical sample number of line learning sample adds 1；The sample number of other random fern numerical value it is constant (in the present embodiment,
Classification is ckTotal sample number M add 1, random fern FlNumerical value be that 5 sample number N plus 1, the sample number N of other numerical valueotherNo
Become)；
6.3) according to the sample number after renewal, Posterior probability distribution of the random fern numerical value in the sample class is recalculated
(in the present embodiment, random fern FlNumerical value be that 5 posterior probability is changed intoThe posterior probability values of other numerical value are changed into)；
6.4) an on-line study sample is often increased newly, just repeatedly 6.1) to 6.3) being updated one to Posterior probability distribution
It is secondary.
Tested by field of traffic, (in realistic objective detection process, we use several differences as shown in Figure 3
Yardstick carries out target detection in video image, and the corresponding frames images of different scale are of different sizes, therefore can detect i.e. frame
Select different size of image block), wherein Fig. 3 a-3d are testing results before on-line study (i.e. only by the inspection of initial training
Survey result), Fig. 3 e-3h are the testing results after on-line study, it can be found that initial training grader is to target detection from figure
Effect it is relatively low, it is high much by the effect after training to target detection.
Fig. 4 is the grader autonomous learning procedure chart under night illumination condition, and wherein Fig. 4 (a) -4 (d) is the beginning of video
In the stage, it is seen that missing inspection is more, this is caused because complete autonomous on-line training positive sample is less.With the increasing of on-line training sample
Many, verification and measurement ratio increases, and false-alarm also progressively increases, such as shown in Fig. 4 (e) -4 (h).After grader further on-line study, its is every
The posterior probability of individual random fern is tended towards stability, and the vehicle target for detecting is also tended to accurately, such as shown in Fig. 4 (i) -4 (l).
Fig. 5 is the grader autonomous learning procedure chart of pedestrian detection, and wherein Fig. 5 (a) -5 (d) is at the beginning of complete autonomous on-line study
The detection case of phase, Fig. 5 (e) -5 (h) is the target detection situation after 200 frames of system autonomous learning, from figure it can be found that
Complete autonomous on-line study method can gradually step up target detection performance.
Fig. 6 compares figure for one embodiment of the invention with the ROC curve of other classical on-line study processes, can be sent out from figure
Now complete autonomous on-line study method has preferable Detection results.
Claims (5)
1. a kind of complete autonomous on-line study method based on random fern grader, it is characterised in that：It comprises the following steps：
1) sample set of initial training grader is prepared：
For frame of video to be detected, frame selects a Target Photo in the video frame, and affine transformation is carried out to the Target Photo
The picture for obtaining is used as positive sample；Not contain the background image region of target as negative sample；So random acquisition is certain
The positive sample and negative sample of quantity as initial training grader sample set；Positive and negative samples are size identical image block；
2) random fern grader initial training：
Initial training is carried out to random fern grader using the sample set of ready initial training grader, positive negative sample is counted
Posterior probability distribution in each random fern；
3) the good random fern grader of initial training is traveled through into frame of video to be detected as current goal detector carries out target
Detection, obtains object module, and calculate the confidence level of each object module；
4) positive and negative sample form collection is built：
Following three kinds of samples are added to positive sample template set M as positive sample template+, remaining is added to negative sample template set M-：
A, step 1) in the positive sample that obtains；
B, to step 3) in the confidence level that obtains exceed the object module of confidence level preset value, using optical flow method to video where it
Frame is tracked and obtains tracking module, if tracking module has overlapping region with the object module, and coincidence factor exceedes default coincidence
Rate, then it is assumed that the tracking module is real goal, M is added to as positive sample template+In；
C, to step 3) in object module of the confidence level more than 0.6 that obtains, using optical flow method frame of video where it is carried out with
Track obtains tracking module, if tracking module has overlapping region with the object module, and coincidence factor then leads to not less than default coincidence factor
Cross conservative similarity ScJudge that can the tracking module add positive sample template set：
Wherein：
If ScMore than default conservative similarity threshold, then the tracking module is used as positive sample template addition M+,To treat point
The similarity of class sample and the first half template of current positive sample template set, S+、S-Sample respectively to be sorted and positive and negative sample
The similarity of this template set,It is two similarities of picture frame, p+, p-Respectively positive sample and negative sample, p is to treat point
Class sample, sample to be sorted is tracking module in this step；
A positive sample template is often added, then the image block for taking in same frame of video four formed objects around it determines whether
Negative sample, if adding negative sample template set M as negative sample template-；
5) nearest neighbor classifier is used, the positive negative sample of on-line study is obtained：
The setting of nearest neighbor classifier is as follows：For each sample p to be sorted, its phase with positive and negative sample form collection is calculated respectively
Like degree S+(p,M+) and S-(p,M-)：
Similarity S can be obtained accordinglyr：
If similarity SrMore than threshold θNN, then judge that the sample to be sorted is real goal, as the positive sample of on-line study；It is no
It is then false-alarm, as the negative sample of on-line study；
Sample to be sorted is step 3 in this step) object module and step 4 that obtain) the positive and negative sample form collection that obtains；
(6) on-line training of random fern grader：
Using step 5) obtain on-line study positive negative sample, on-line study is carried out to random fern grader, gradually step up it
Nicety of grading；
Target detection is carried out using the random fern grader of on-line study as the detecting system of sustainable renewal；
Described step 4) also include that template set cuts down mechanism：Sample to be sorted is equal to be sorted with the similarity of positive and negative template set
In sample and positive and negative template set between single positive and negative sample form similarity maximum；Real-time statistics each positive and negative sample forms
The number of times of the maximum is obtained, if the number of times of the maximum of certain positive and negative sample form acquisition is less than maximum number of times preset value,
Then remove corresponding positive sample template or negative sample template.
2. the complete autonomous on-line study method based on random fern grader according to claim 1, it is characterised in that：It is described
The step of 2) specific method it is as follows：
2.1) random fern is constructed：
To taking s to characteristic point as one group of random fern at random in the single sample in the sample set of initial training grader, each
The position that sample takes characteristic point is identical, and each pair characteristic point carries out the comparing of pixel value, previous characteristic point picture in each pair characteristic point
The big characteristic value that then takes of plain value is 1, otherwise it is s characteristic value being obtained more afterwards to characteristic point of 0, s according to random then to take characteristic value
Order constitute the binary number of s, as this group random fern numerical value of random fern, feature in the random fern of each sample
The sequence consensus of value；
2.2) posterior probability of the random fern numerical value in positive negative sample class is calculated：
In random fern, some is obtained for positive sample, and other are obtained for negative sample；The value species of random fern numerical value has
2sIt is individual；
The positive sample number of the value of every kind of random fern numerical value is counted, so as to obtain random fern numerical value in positive sample class C1On after
Test probability distribution P (Fl|C1)；Random fern numerical value is similarly obtained in negative sample class C0On Posterior probability distribution P (Fl|C0)；Joint
All random ferns are classified to the sample set of initial training grader, as random fern grader；
Described step 3) target detection is carried out in every frame video image using above-mentioned random fern grader：
Traversal every frame video image to be detected, the image block of extraction formed objects is used as treating test sample in every frame video image
This, size and the step 1 of sample to be tested) in positive sample it is equal in magnitude, calculate the random fern numerical value of each sample to be tested so that
Corresponding posterior probability is obtained, finally by its classification of random fern classifier calculated；
It is the image block of positive sample for classification, then is detected as target.
3. the complete autonomous on-line study method based on random fern grader according to claim 1, it is characterised in that：It is described
The step of 4) often add a positive sample template, then taking in same frame of video four image blocks of formed objects around it and judging is
It is no for negative sample when, introduce Gaussian Background modeling, if in image block foreground pixel be less than foreground pixel threshold value, judge that it is negative
Sample.
4. the complete autonomous on-line study method based on random fern grader according to claim 2, it is characterised in that：It is described
The step of 6) on-line study of random fern grader realized by updating Posterior probability distribution.
5. the complete autonomous on-line study method based on random fern grader according to claim 4, it is characterised in that：It is described
The step of 6) specific method is as follows：
6.1) using step 5) the positive negative sample that obtains is used as on-line study sample；If an on-line study sample is (fnew, ck), its
Middle fnewIt is the binary number of random fern s, ckIt is sample class, calculates the random fern numerical value of the on-line study sample；
6.2) to step 2.1) classification is c in sample setkTotal sample number add 1, classification is ckWith the on-line study sample with
Machine fern numerical value identical sample number adds 1；The sample number of other random fern numerical value is constant；
6.3) according to the sample number after renewal, Posterior probability distribution of the random fern numerical value in the sample class is recalculated；
6.4) an on-line study sample is often increased newly, just repeatedly 6.1) to 6.3) being updated once to Posterior probability distribution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410407669.2A CN104156734B (en) | 2014-08-19 | 2014-08-19 | A kind of complete autonomous on-line study method based on random fern grader |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410407669.2A CN104156734B (en) | 2014-08-19 | 2014-08-19 | A kind of complete autonomous on-line study method based on random fern grader |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104156734A CN104156734A (en) | 2014-11-19 |
| CN104156734B true CN104156734B (en) | 2017-06-13 |
Family
ID=51882231
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410407669.2A Expired - Fee Related CN104156734B (en) | 2014-08-19 | 2014-08-19 | A kind of complete autonomous on-line study method based on random fern grader |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104156734B (en) |
Families Citing this family (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105825233B (en) * | 2016-03-16 | 2019-03-01 | 中国地质大学（武汉） | A kind of pedestrian detection method based on on-line study random fern classifier |
| CN107292918B (en) * | 2016-10-31 | 2020-06-19 | 清华大学深圳研究生院 | Tracking method and device based on video online learning |
| CN106846362B (en) * | 2016-12-26 | 2020-07-24 | 歌尔科技有限公司 | Target detection tracking method and device |
| CN106845387B (en) * | 2017-01-18 | 2020-04-24 | 合肥师范学院 | Pedestrian detection method based on self-learning |
| CN107092878A (en) * | 2017-04-13 | 2017-08-25 | 中国地质大学（武汉） | It is a kind of based on hybrid classifer can autonomous learning multi-target detection method |
| CN108932857B (en) * | 2017-05-27 | 2021-07-27 | 西门子（中国）有限公司 | Method and device for controlling traffic signal lamp |
| CN107423702B (en) * | 2017-07-20 | 2020-06-23 | 西安电子科技大学 | Video target tracking method based on TLD tracking system |
| CN108038515A (en) * | 2017-12-27 | 2018-05-15 | 中国地质大学（武汉） | Unsupervised multi-target detection tracking and its storage device and camera device |
| CN109325966B (en) * | 2018-09-05 | 2022-06-03 | 华侨大学 | Method for carrying out visual tracking through space-time context |
| CN110135456A (en) * | 2019-04-08 | 2019-08-16 | 图麟信息科技（上海）有限公司 | A kind of training method and device of target detection model |
| CN111861966B (en) * | 2019-04-18 | 2023-10-27 | 杭州海康威视数字技术股份有限公司 | Model training method and device and defect detection method and device |
| CN110211153A (en) * | 2019-05-28 | 2019-09-06 | 浙江大华技术股份有限公司 | Method for tracking target, target tracker and computer storage medium |
| CN110717556B (en) * | 2019-09-25 | 2023-04-07 | 南京旷云科技有限公司 | Posterior probability adjusting method and device for target recognition |
| CN110889747B (en) * | 2019-12-02 | 2023-05-09 | 腾讯科技（深圳）有限公司 | Commodity recommendation method, device, system, computer equipment and storage medium |
| CN113066101A (en) * | 2019-12-30 | 2021-07-02 | 阿里巴巴集团控股有限公司 | Data processing method and device, and image processing method and device |
| CN112257738A (en) * | 2020-07-31 | 2021-01-22 | 北京京东尚科信息技术有限公司 | Training method and device of machine learning model and classification method and device of image |
| CN112347968A (en) * | 2020-11-18 | 2021-02-09 | 合肥湛达智能科技有限公司 | Target detection method based on autonomous online learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102982340A (en) * | 2012-10-31 | 2013-03-20 | 中国科学院长春光学精密机械与物理研究所 | Target tracking method based on semi-supervised learning and random fern classifier |
| CN103208190A (en) * | 2013-03-29 | 2013-07-17 | 西南交通大学 | Traffic flow detection method based on object detection |
| CN103699908A (en) * | 2014-01-14 | 2014-04-02 | 上海交通大学 | Joint reasoning-based video multi-target tracking method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101332630B1 (en) * | 2012-06-18 | 2013-11-25 | 한국과학기술원 | Weight lightened random ferns and image expression method using the same |
-
2014
- 2014-08-19 CN CN201410407669.2A patent/CN104156734B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102982340A (en) * | 2012-10-31 | 2013-03-20 | 中国科学院长春光学精密机械与物理研究所 | Target tracking method based on semi-supervised learning and random fern classifier |
| CN103208190A (en) * | 2013-03-29 | 2013-07-17 | 西南交通大学 | Traffic flow detection method based on object detection |
| CN103699908A (en) * | 2014-01-14 | 2014-04-02 | 上海交通大学 | Joint reasoning-based video multi-target tracking method |
Non-Patent Citations (1)
| Title |
|---|
| "基于自主学习的复杂目标跟踪算法研究";徐诚;《中国优秀硕士学位论文全文数据库 信息科技辑》;20131215(第S1期);第35-37、49-54页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104156734A (en) | 2014-11-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104156734B (en) | A kind of complete autonomous on-line study method based on random fern grader | |
| CN108830188B (en) | Vehicle detection method based on deep learning | |
| CN109977812B (en) | Vehicle-mounted video target detection method based on deep learning | |
| CN111931684B (en) | Weak and small target detection method based on video satellite data identification features | |
| CN104063713B (en) | A kind of semi-autonomous on-line study method based on random fern grader | |
| CN107609525B (en) | Remote sensing image target detection method for constructing convolutional neural network based on pruning strategy | |
| CN105975941B (en) | A kind of multi-direction vehicle detection identifying system based on deep learning | |
| CN106897738B (en) | A kind of pedestrian detection method based on semi-supervised learning | |
| CN101814149B (en) | Self-adaptive cascade classifier training method based on online learning | |
| CN109767427A (en) | The detection method of train rail fastener defect | |
| CN110084165A (en) | The intelligent recognition and method for early warning of anomalous event under the open scene of power domain based on edge calculations | |
| CN104850865B (en) | A kind of Real Time Compression tracking of multiple features transfer learning | |
| CN111310756B (en) | Damaged corn particle detection and classification method based on deep learning | |
| CN104463196A (en) | Video-based weather phenomenon recognition method | |
| CN107506703A (en) | A kind of pedestrian's recognition methods again for learning and reordering based on unsupervised Local Metric | |
| CN107194418A (en) | A kind of Aphids in Rice Field detection method based on confrontation feature learning | |
| CN110263712A (en) | A kind of coarse-fine pedestrian detection method based on region candidate | |
| CN108596038A (en) | Erythrocyte Recognition method in the excrement with neural network is cut in a kind of combining form credit | |
| CN110929746A (en) | Electronic file title positioning, extracting and classifying method based on deep neural network | |
| CN112926522B (en) | Behavior recognition method based on skeleton gesture and space-time diagram convolution network | |
| CN108256462A (en) | A kind of demographic method in market monitor video | |
| CN105976397A (en) | Target tracking method based on half nonnegative optimization integration learning | |
| CN105427339A (en) | Characteristic screening and secondary positioning combined fast compression tracking method | |
| CN111144462A (en) | Unknown individual identification method and device for radar signals | |
| CN106548195A (en) | A kind of object detection method based on modified model HOG ULBP feature operators |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170613 Termination date: 20180819 |