CN104156734A - Fully-autonomous on-line study method based on random fern classifier - Google Patents

Abstract

The invention discloses a fully autonomous online learning method based on a random fern classifier. The method only needs to select a target in a video frame once to perform online learning of the classifier for the target class. The steps are: first, use affine transformation to obtain the initial positive sample set for the frame-selected target, and extract a small number of negative sample sets in the non-target area of the video to train the initial random fern classifier; secondly, use this classifier to perform Target Detection. During the detection process, the nearest neighbor classifier is used to collect online learning new samples, and automatically judge the sample category; finally, the new samples are used for online training of the random fern classifier, and the posterior probability of the random fern is updated to gradually improve the target detection of the classifier. Accuracy, to achieve fully autonomous online learning of the target detection system.

Description

A Fully Autonomous Online Learning Method Based on Random Fern Classifier

technical field

The invention belongs to the field of pattern recognition, and in particular relates to a fully autonomous online learning method based on a random fern classifier.

Background technique

Online learning belongs to the research category of incremental learning. In this type of method, the classifier only learns each sample once instead of repeated learning. In this way, a large amount of storage space is not required to store training samples during the operation of the online learning algorithm. Every time the classifier obtains a sample, it conducts online learning. Through online learning, the classifier can still update and improve itself according to the new sample during use, and further improve the classification effect.

Early online learning algorithms include Winnow algorithm, unified linear prediction algorithm, etc. In 2001, scholar Oza combined these algorithms with boosting algorithm and proposed online boosting algorithm (this algorithm is quoted from "Online bagging and boosting" N.Oza and S.Russell , In Proc.Artificial Intelligence and Statistics,105-112,2001), in Oza's method, each feature corresponds to a weak classifier, and the strong classifier is a weighted sum of a certain number of weak classifiers, where the weak classifier are selected from the set of weak classifiers. During online learning, each training sample updates each weak classifier in the weak classifier set one by one, including adjusting the classification threshold of positive and negative samples and the weight of the classifier, so that the weight of a good weak classifier is getting higher and higher. And the weight of the poorer weak classifier is getting lower and lower, so that each time a sample is learned online, a weak classifier with the highest current weight can be selected and added to the strong classifier, so that the final trained classifier has a stronger classification ability. .

However, each weak classifier in the weak classifier set of the online boosting algorithm needs to learn new samples online. When the number of weak classifiers is large, the online learning speed will inevitably slow down. Grabner has improved the online boosting algorithm so that it can also perform feature selection like the Adaboost algorithm, and this feature selection and update of the classifier are all performed online, which is called online Adaboost (the algorithm is quoted from "On- line boosting and vision" H. Grabner and H. Bischof, In Proc. CVPR, (1): 260-267, 2006). However, online Adaboost uses feature selection operators instead of general weak classifiers to synthesize strong classifiers. The number of feature selection operators and the number of weak classifiers corresponding to feature selection operators are fixed, and the corresponding online learning classifier structure is relatively rigid. When it is found that its classification ability cannot meet the requirements of detection performance, even continuous online learning cannot improve the detection accuracy.

Ozuysal no longer uses weak classifiers to form strong classifiers for sample classification, but randomly extracts multiple features from the sample feature set to form a random fern, and trains the posterior probability distribution of samples through random fern statistics, and then uses the posterior probability distribution of multiple random ferns The random fern classifier algorithm (this algorithm is quoted from "Fast keypoint recognition using random ferns" In Pattern Analysis Machine Intelligence, IEEE Transaction on 32(3), 448-461, 2010).

Contents of the invention

The technical problem to be solved by the present invention is to provide a fully autonomous online learning method based on a random fern classifier, which is used for autonomous learning of classifiers to improve classification performance.

The technical scheme that the present invention takes for solving the above-mentioned technical problem is: a kind of fully autonomous online learning method based on random fern classifier, it is characterized in that: it comprises the following steps:

1) Prepare the sample set for the initial training classifier:

For the video frame to be detected, a target picture is selected in the video frame, and the picture obtained by affine transformation of the target picture is used as a positive sample; the background image area that does not contain the target is used as a negative sample; in this way, a certain The number of positive samples and negative samples is used as the sample set for the initial training classifier; the positive and negative samples are image blocks of the same size;

2) Random fern classifier initial training:

Use the sample set of the prepared initial training classifier to perform initial training on the random fern classifier, and count the posterior probability distribution of positive and negative samples on each random fern;

3) Use the initially trained random fern classifier as the current target detector to traverse the video frames to be detected for target detection, obtain target modules, and calculate the confidence of each target module;

4) Construct positive and negative sample template sets:

Add the following three samples as positive sample templates to the positive sample template set M ⁺ , and add the rest to the negative sample template set M ⁻ :

A, the positive sample obtained in step 1);

B. For the target module whose confidence degree obtained in step 3) exceeds the preset value of the confidence degree, the optical flow method is used to track the video frame where it is located to obtain the tracking module. If the tracking module has an overlapping area with the target module, and the coincidence rate If it exceeds the preset coincidence rate, the tracking module is considered to be a real target and added to M ⁺ as a positive sample template;

C. For the target module whose confidence degree obtained in step 3) exceeds the preset value of the confidence degree, the optical flow method is used to track the video frame where it is located to obtain the tracking module. If the tracking module has an overlapping area with the target module, and the coincidence rate If the preset coincidence rate is not exceeded, it is judged whether the tracking module can be added to the positive sample template set by the conservative similarity S ^c :

${\mathrm{<span\; class="notranslate">\; S</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; 50</span>}<span\; class="notranslate">\; \%</span>}^{<span\; class="notranslate">\; +</span>}}{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; 50</span>}<span\; class="notranslate">\; \%</span>}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; -</span>}}$

in:

If S ^c is greater than the preset conservative similarity threshold, the tracking module is added to M ⁺ as a positive sample template, is the similarity between the sample to be classified and the first half template of the current positive sample template set, S ⁺ and S ^- are the similarity between the sample to be classified and the positive and negative sample template sets respectively, Be the similarity of two image frames, p ⁺ , p ^- are positive samples and negative samples respectively, p is the sample to be classified, and the sample to be classified is the tracking module in this step;

Add a positive sample template every time, then get four image blocks of the same size around it in the same video frame to judge whether it is a negative sample, if add the negative sample template set M as a negative sample template ⁻ ;

5) Use the nearest neighbor classifier to obtain positive and negative samples for online learning:

The setting of the nearest neighbor classifier is as follows: For each sample p to be classified, calculate its similarity S ⁺ (p, M ⁺ ) and S ^- (p, M ^- ) with the template set of positive and negative samples respectively:

${\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; max</span>}}_{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; +</span>}}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; max</span>}}_{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; -</span>}}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span>$

The corresponding available similarity S ^r :

${\mathrm{<span\; class="notranslate">\; S</span>}}^{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; +</span>}}{{\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; -</span>}}$

If the similarity S ^r is greater than the threshold θ _NN , it is judged that the sample to be classified is a real target, and it is used as a positive sample for online learning; otherwise, it is a false alarm, and it is used as a negative sample for online learning;

In this step, the sample to be classified is the target module obtained in step 3) and the positive and negative sample template set obtained in step 4);

(6) Online training of random fern classifier:

Use the positive and negative samples of the online learning that step 5) obtains, carry out online learning to random fern classifier, improve its classification accuracy gradually;

Using an Online-Learned Random Fern Classifier as a Continuously Renewable Detection System for Object Detection.

According to the above scheme, the concrete method of described step 2) is as follows:

2.1) Construct a random fern:

Randomly select s pairs of feature points on a single sample in the sample set of the initial training classifier as a group of random ferns. The position of each feature point is the same for each sample, and the pixel value of each pair of feature points is compared. The top of each pair of feature points If the pixel value of a feature point is large, the feature value is taken as 1, otherwise, the feature value is taken as 0, and the s feature values obtained after comparing the feature points with s form a binary number of s bits in a random order, which is the set of random The random fern value of the fern, the order of the eigenvalues in the random fern of each sample is the same;

2.2) Calculate the posterior probability of random fern values on positive and negative sample classes:

Among the random ferns, some are obtained from positive samples, and others are obtained from negative samples; there are 2 ^s types of random fern values;

Count the number of positive samples of each random fern value, so as to obtain the posterior probability distribution P(F _l |C ₁ ) of the random fern value on the positive sample class C ₁ ; similarly, obtain the random fern value on the negative sample Posterior probability distribution P(F _l |C ₀ ) on class C ₀ ; combine all random ferns to classify the sample set of the initial training classifier, which is a random fern classifier;

Described step 3) adopt above-mentioned random fern classifier to carry out target detection in each frame video image:

Traversing each frame of video image to be detected, extracting an image block of the same size in each frame of video image as the sample to be tested, the size of the sample to be tested is equal to the size of the positive sample in step 1), and calculating the random fern of each sample to be tested value, so as to obtain the corresponding posterior probability, and finally calculate its category by the random fern classifier;

For the image block whose category is a positive sample, it is detected as an object.

According to the above-mentioned scheme, described step 4) each time a positive sample template is added, four image blocks of the same size around it in the same video frame are taken to judge whether they are negative samples, and Gaussian background modeling is introduced. If the foreground in the image block If the pixel is smaller than the foreground pixel threshold, it is judged as a negative sample.

According to the above scheme, described step 4) also includes a template set reduction mechanism: the similarity between the sample to be classified and the positive and negative template set is equal to the maximum value of the similarity between the sample to be classified and the single positive and negative sample template in the positive and negative template set; Count the number of times each positive and negative sample template obtains the maximum value in real time, and if the number of times that a certain positive and negative sample template obtains the maximum value is less than the preset value of the maximum number of times, then remove the corresponding positive sample template or negative sample template.

According to the above scheme, the step 6) online learning of the random fern classifier is realized by updating the posterior probability distribution.

According to the above-mentioned scheme, described step 6) concrete method is as follows:

6.1) Use the positive and negative samples obtained in step 5) as an online learning sample; set an online learning sample as (f _new , c _k ), where f _new is a random binary number of s bits, c _k is a sample category, and calculate the The random fern value of online learning samples;

6.2) Add 1 to the total number of samples whose category is c _k in the sample set in step 2.1), and add 1 to the number of samples whose category is c _k and the random fern value of the online learning sample; the number of samples of other random fern values remains unchanged;

6.3) According to the updated number of samples, recalculate the posterior probability distribution of the random fern value on the sample class;

6.4) For each new online learning sample, repeat 6.1) to 6.3) to update the posterior probability distribution once.

The beneficial effects of the present invention are:

1. Only need to select a target in the video frame once to carry out online learning of the classifier for this target class, that is: first, use affine transformation to obtain the initial positive sample set for the frame-selected target, and in the non-target area of the video Extract a small number of negative sample sets to train the initial random fern classifier; secondly, use the classifier to detect objects in video frames; during the detection process, use the nearest neighbor classifier to collect online learning new samples, and automatically judge the sample category; finally , the online learning new samples are used for the online training of the random fern classifier, the posterior probability of the random fern is updated, the accuracy of the target detection of the random fern classifier is gradually improved, and the target detection system is fully autonomous online learning.

2. This patent introduces a template set reduction mechanism, which can avoid the disadvantages of template concentration and a decrease in system operating speed that may be caused by more positive and negative sample templates.

Description of drawings

Fig. 1 is the flowchart of the inventive method;

Fig. 2 is a structural diagram of a random fern classifier in an embodiment of the present invention;

Fig. 3 is the contrast figure of detection performance before and after the random fern classifier of online learning of an embodiment of the present invention, wherein Fig. 3 (a)-3 (d) is the detection result before online learning, Fig. 3 (i)-3 (l ) is the detection result after online learning;

Figure 4 is a diagram of the autonomous learning process of the classifier under night light conditions;

Fig. 5 is a self-learning process diagram of a classifier for pedestrian detection;

Fig. 6 is a comparison chart of ROC curves between an embodiment of the present invention and other classic online learning processes.

Detailed ways

The present invention will be further described below in conjunction with specific examples and accompanying drawings.

The invention discloses a nearest neighbor classifier training method in a fully autonomous online learning process based on the research of a target detection system. The method only needs to select a target in a video frame once to perform online learning of the classifier for the target class. The steps are: first, use affine transformation to obtain the initial positive sample set for the frame-selected target, and extract a small number of negative sample sets in the non-target area of the video to train the initial random fern classifier; secondly, use this classifier to perform Target Detection. During the detection process, the nearest neighbor classifier is used to collect online learning new samples, and automatically judge the sample category; finally, the new samples are used for online training of the random fern classifier, and the posterior probability of the random fern is updated to gradually improve the target detection of the classifier. Accuracy, to achieve fully autonomous online learning of the target detection system.

The present invention provides a fully autonomous online learning method based on a random fern classifier as shown in Figure 1, comprising the following steps:

1) Prepare the sample set for the initial training classifier:

For the video frame to be detected, a target is selected in the first frame of the video image, and the image obtained by affine transformation of the target image is used as a positive sample; the background image area that does not contain the target is used as a negative sample; so random Obtain a certain number of positive samples and negative samples as the sample set for initial training classifier.

In this embodiment, the samples in the sample set of the initial training classifier are image blocks of the same size, generally 15×15 (pixels), if the image block contains a target to be detected, then the sample is a positive sample, and there is no is a negative sample.

2) Random fern classifier initial training:

Use the sample set of the prepared initial training classifier to perform initial training on the random fern classifier, and count the posterior probability distribution of positive and negative samples on each random fern, as shown in Figure 2.

The specific method is as follows:

2.1) Construct a random fern:

Randomly get s pairs of feature points on a single sample in the sample set as a group of random ferns (5 pairs are selected in this embodiment), the position of each feature point is the same for each sample, and the comparison of pixel values is carried out for each pair of feature points, and each pair of feature points If the pixel value of the previous feature point in the point is large, then the feature value is 1, otherwise, the feature value is 0, and the s feature values obtained after comparing the feature points in s form a binary number of s bits in random order, that is, The random fern value of this group of random ferns, the order of the eigenvalues in the random ferns of each sample is the same;

2.2) Calculate the posterior probability of random fern values on positive and negative sample classes:

Among the random ferns, some are obtained from positive samples, and others are obtained from negative samples; the features contained in the random fern F _l of each sample can be combined to form a decimal number, because the decimal number is obtained by S-bit binary code, Therefore, there are 2 ^s value types of the random fern value, that is, 2 ^s possibilities ( ²⁵ possibilities in this embodiment);

Count the number of positive samples of each random fern value, so as to obtain the posterior probability distribution P(F _l |C ₁ ) of the random fern value on the positive sample class C ₁ ; similarly, obtain the random fern value on the negative sample Posterior probability distribution P(F _l |C ₀ ) on class C ₀ ; all random ferns are combined to classify the sample set of the initial training classifier, which is a random fern classifier.

3) Use the initially trained random fern classifier as the current target detector to traverse the video frames to be detected for target detection, obtain the target modules, and calculate the confidence of each target module, specifically: traverse the video frames to be detected, Extract an image block of the same size in the video frame as the sample to be tested. The size of the sample to be tested is equal to the size of the positive sample in step 1), and the random fern value of each sample to be tested is calculated to obtain the corresponding posterior probability. Finally its class is computed by a random fern classifier;

For the image block whose category is a positive sample, it is detected as an object and becomes an object module.

4) Construct positive and negative sample template sets:

Add the following three samples as positive sample templates to the positive sample template set M ⁺ , and add the rest to the negative sample template set M ⁻ :

A, the positive sample obtained in step 1);

B. For the target module whose confidence degree obtained in step 3) exceeds the preset value of confidence degree (preferably 0.6), use the optical flow method to track the video frame where it is located to obtain the tracking module. If the tracking module has an overlapping area with the target module , and the coincidence rate exceeds the preset coincidence rate (the preset coincidence rate is usually 60%), then the tracking module is considered to be a real target and added to M ⁺ as a positive sample template;

C. For the target module whose confidence degree obtained in step 3) exceeds the preset value of the confidence degree (preferably 0.6), use the optical flow method to track the video frame where it is located to obtain the tracking module. If the tracking module has an overlapping area with the target module , and the coincidence rate does not exceed the preset coincidence rate, then judge whether the tracking module can be added to the positive sample template set by conservative similarity S ^c :

${\mathrm{<span\; class="notranslate">\; S</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; 50</span>}<span\; class="notranslate">\; \%</span>}^{<span\; class="notranslate">\; +</span>}}{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; 50</span>}<span\; class="notranslate">\; \%</span>}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; -</span>}}$

in:

If S ^c is greater than the preset conservative similarity threshold (preferably 0.6), the tracking module is added to M ⁺ as a positive sample template, is the similarity between the sample to be classified and the first half template of the current positive sample template set, S ⁺ and S ^- are the similarity between the sample to be classified and the positive and negative sample template sets respectively, Be the similarity of two image frames, p ⁺ , p ^- are positive samples and negative samples respectively, p is the sample to be classified, and the sample to be classified is the tracking module in this step;

Each time a positive sample template is added, four surrounding image blocks of the same size in the same video frame are taken to judge whether it is a negative sample, and if it is used as a negative sample template, it is added to the negative sample template set M ⁻ . When judging, Gaussian background modeling is introduced. If the foreground pixel in the image block is smaller than the foreground pixel threshold (preferably less than 30%), it is judged as a negative sample.

Step 4) also includes a template set reduction mechanism: the similarity between the sample to be classified and the positive and negative template set is equal to the maximum similarity between the sample to be classified and a single positive and negative sample template in the positive and negative template set; real-time statistics of each positive and negative sample template The number of times the maximum value is obtained. If the number of times the maximum value is obtained by a certain positive and negative sample template is less than the preset value of the maximum number of times, the corresponding positive sample template or negative sample template is removed.

5) Use the nearest neighbor classifier to obtain positive and negative samples for online learning:

The setting of the nearest neighbor classifier is as follows: For each sample p to be classified, calculate its similarity S ⁺ (p, M ⁺ ) and S ^- (p, M ^- ) with the template set of positive and negative samples respectively:

${\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; max</span>}}_{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; +</span>}}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; max</span>}}_{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; -</span>}}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span>$

The corresponding available similarity S ^r :

${\mathrm{<span\; class="notranslate">\; S</span>}}^{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; +</span>}}{{\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; S</span>}}^{<span\; class="notranslate">\; -</span>}}$

If the similarity S ^r is greater than the threshold θ _NN , it is judged that the sample to be classified is a real target, and it is used as a positive sample for online learning; otherwise, it is a false alarm, and it is used as a negative sample for online learning;

The samples to be classified in this step are the target modules obtained in step 3) and the positive and negative sample template sets obtained in step 4).

(6) Online training of random fern classifier:

Use the positive and negative samples of online learning obtained in step 5) to conduct online learning on the random fern classifier, and gradually improve its classification accuracy; use the online learning random fern classifier as a continuously updated detection system for target detection.

The online learning of the random fern classifier is realized by updating the posterior probability distribution, and the specific method is as follows:

6.1) The positive and negative samples obtained in step 5) are used as online learning samples; an online learning sample is (f _new , c _k ), where f _new is a random binary number of s bits (in this embodiment, f _new is 00101 , that is, the decimal number 5), c _k is the sample category, and calculates the random fern value of the online learning sample;

6.2) As shown in Figure 2, add 1 to the total number of samples whose category is c _k in the sample set in step 2.1), add 1 to the number of samples whose category is c _k and the same random fern value of the online learning sample; The number of samples is constant (in the present embodiment, the total number of samples M of the category c _k adds 1, the value of the random fern F _l is 5 and the number of samples N adds 1, and the number of samples N _other of other values is constant);

6.3) According to the updated number of samples, recalculate the posterior probability distribution of the random fern value on the sample class (in this embodiment, the posterior probability that the value of the random fern _F1 is 5 becomes The posterior probability values for other values become );

6.4) For each new online learning sample, repeat 6.1) to 6.3) to update the posterior probability distribution once.

Through experiments in the traffic field, as shown in Figure 3 (in the actual target detection process, we use several different scales for target detection in video images, and the image frames corresponding to different scales have different sizes, so it can be detected and framed out Image blocks of different sizes), where Figure 3a-3d is the detection result before online learning (that is, the detection result only through initial training), and Figure 3e-3h is the detection result after online learning, from which it can be found that the initial training classification The effect of the device on target detection is low, and the effect on target detection is much higher after training.

Figure 4 is a diagram of the autonomous learning process of the classifier under nighttime lighting conditions, in which Figures 4(a)-4(d) are the beginning stages of the video, and it can be seen that there are many missed detections, which is due to the lack of positive samples for fully autonomous online training of. With the increase of online training samples, the detection rate increases, and the false alarms gradually increase, as shown in Figure 4(e)-4(h). When the classifier is further studied online, the posterior probability of each random fern tends to be stable, and the detected vehicle targets also tend to be accurate, as shown in Figure 4(i)-4(l).

Figure 5 is a diagram of the autonomous learning process of the classifier for pedestrian detection, in which Figures 5(a)-5(d) show the detection situation at the initial stage of fully autonomous online learning, and Figure 5(e)-5(h) show that the system has learned 200 times independently The target detection situation after the frame. From the figure, it can be found that the fully autonomous online learning method can gradually improve the target detection performance.

FIG. 6 is a comparison diagram of ROC curves between an embodiment of the present invention and other classic online learning processes. From the figure, it can be found that the fully autonomous online learning method has a better detection effect.

Claims (6)

1. the complete autonomous on-line study method based on random fern sorter, is characterized in that: it comprises the steps:

1) prepare the sample set of initial training sorter:

For frame of video to be detected, at frame of video center, select a Target Photo, this Target Photo is carried out to picture that affined transformation obtains as positive sample; Using do not contain target background image region as negative sample; The so random positive sample that obtains some and negative sample are as the sample set of initial training sorter; Positive and negative samples is the image block that size is identical;

2) random fern sorter initial training:

Use the sample set of ready initial training sorter to carry out initial training to random fern sorter, add up the posterior probability of positive negative sample on each random fern and distribute;

3) the good random fern sorter of initial training is traveled through to frame of video to be detected as current goal detecting device and carry out target detection, obtain object module, and calculate the degree of confidence of each object module;

4) build positive negative sample template set:

Using following three kinds of samples as positive sample template, add positive sample template set M to ⁺, all the other add negative sample template set M to ^-:

A, step 1) in the positive sample that obtains;

B, to step 3) in the degree of confidence that obtains surpass the object module of degree of confidence preset value, adopt optical flow method to follow the tracks of and obtain tracking module its place frame of video, if tracking module and this object module have overlapping region, and coincidence factor surpasses default coincidence factor, think that this tracking module is real goal, as positive sample template, add M to ⁺in;

C, to step 3) in the degree of confidence that obtains surpass 0.6 object module, adopt optical flow method to follow the tracks of and obtain tracking module its place frame of video, if tracking module and this object module have overlapping region, and coincidence factor is over default coincidence factor, by conservative similarity S ^cjudge that can this tracking module add positive sample template set:

$S^c=\frac{S_{50\%}^{+}}{S_{50\%}^{+}+S^{-}}$

Wherein:

If S ^cbe greater than default conservative similarity threshold, this tracking module adds M as positive sample template ⁺, for the similarity of the first half template of sample to be sorted and current positive sample template set, S ⁺, S ^-be respectively the similarity of sample to be sorted and positive and negative samples template set, be the similarity of two picture frames, p ⁺, p ^-be respectively positive sample and negative sample, p is sample to be sorted, and in this step, sample to be sorted is tracking module;

Often add a positive sample template, get in same frame of video its around the image block of four formed objects determine whether negative sample, if add negative sample template set M as negative sample template ^-;

5) use nearest neighbor classifier, obtain the positive negative sample of on-line study:

Arranging of nearest neighbor classifier is as follows: for each sample p to be sorted, calculate respectively itself and the similarity S of positive negative sample template set ⁺(p, M ⁺) and S ^-(p, M ^-):

$S^{+}(p,M^{+})={\max }_{p_i^{+}\&Element;M^{+}}S(p,p_i^{+})$

$S^{-}(p,M^{-})={\max }_{p_i^{-}\&Element;M^{-}}S(p,p_i^{-})$

Can obtain similarity S accordingly ^r:

$S^r=\frac{S^{+}}{S^{+}+S^{-}}$

If similarity S ^rbe greater than threshold value θ _nN, judge that this sample to be sorted is real goal, as the positive sample of on-line study; Otherwise be false-alarm, as the negative sample of on-line study;

In this step, sample to be sorted is step 3) object module and the step 4 that obtain) the positive negative sample template set that obtains;

(6) the online training of random fern sorter:

The positive negative sample of on-line study use step 5) obtaining, carries out on-line study to random fern sorter, improves gradually its nicety of grading;

The random fern sorter of on-line study is carried out to target detection as the detection system of sustainable renewal.

2. the complete autonomous on-line study method based on random fern sorter according to claim 1, is characterized in that: concrete grammar described step 2) is as follows:

2.1) construct random fern:

To on the single sample in the sample set of initial training sorter, get at random s to unique point as one group of random fern, it is identical that each sample is got the position of unique point, every pair of unique point is carried out the comparison of pixel value, in every pair of unique point, greatly to get eigenwert be 1 to previous unique point pixel value, otherwise get eigenwert, be 0, the s that s obtains more afterwards to a unique point eigenwert forms the binary number of a s position according to random order, be the random fern numerical value that this organizes random fern, the sequence consensus of eigenwert in the random fern of each sample;

2.2) calculate the posterior probability of random fern numerical value in positive negative sample class:

In random fern, some obtains for positive sample, and other obtains for negative sample; The value kind of random fern numerical value has 2 ^sindividual;

Add up the positive number of samples of the value of every kind of random fern numerical value, thereby obtain random fern numerical value at positive sample class C ₁on posterior probability distribution P (F _l| C ₁); In like manner obtain random fern numerical value at negative sample class C ₀on posterior probability distribution P (F _l| C ₀); Combine all random ferns the sample set of initial training sorter is classified, be random fern sorter;

Described step 3) adopt above-mentioned random fern sorter in every frame video image, to carry out target detection:

Travel through every frame video image to be detected, in every frame video image, extract the image block of formed objects as sample to be tested, size and the step 1 of sample to be tested) in positive size equate, calculate the random fern numerical value of each sample to be tested, thereby obtain corresponding posterior probability, finally by random its classification of fern classifier calculated;

The image block that is positive sample for classification, is detected as target.

3. the complete autonomous on-line study method based on random fern sorter according to claim 1, it is characterized in that: described step 4) often add a positive sample template, when getting in same frame of video it around the image block of four formed objects determining whether negative sample, introduce Gaussian Background modeling, if foreground pixel is less than foreground pixel threshold value in image block, judge that it is negative sample.

4. according to the complete autonomous on-line study method based on random fern sorter described in claim 1 or 3, it is characterized in that: described step 4) also comprise that template set subdues mechanism: the similarity of sample to be sorted and positive and negative template set equals in sample to be sorted and positive and negative template set the maximal value of similarity between single positive negative sample template; Each positive negative sample template of real-time statistics obtains this peaked number of times, if this peaked number of times that certain positive negative sample template obtains is less than maximal value number of times preset value, removes corresponding positive sample template or negative sample template.

5. the complete autonomous on-line study method based on random fern sorter according to claim 2, is characterized in that: described step 6) on-line study of random fern sorter is distributed and realized by renewal posterior probability.

6. the complete autonomous on-line study method based on random fern sorter according to claim 5, is characterized in that: described step 6) concrete grammar is as follows:

6.1) using step 5) the positive negative sample that obtains is as on-line study sample; If an on-line study sample is (f _new, c _k), f wherein _newfor the binary number of random fern s position, c _kfor sample class, calculate the random fern numerical value of this on-line study sample;

6.2) to step 2.1) classification is c in sample set _ktotal sample number add 1, classification is c _kthe sample number identical with the random fern numerical value of this on-line study sample add 1; The sample number of other random fern numerical value is constant;

6.3), according to the sample number after upgrading, recalculate the posterior probability of random fern numerical value in this sample class and distribute;

6.4) often increase an on-line study sample newly, just repeat 6.1) to 6.3) posterior probability is distributed and upgraded once.