CN105701469A - Robust population counting method based on cost-sensitive sparse linear regression - Google Patents

Abstract

The invention provides a robust population counting method based on cost-sensitive sparse linear regression. In the method, a supervised learning linearity regression model image population counting method is taken as a basis, and characteristic fusion, a sparse constraint and a cost-sensitive strategy are used to increase robustness and accuracy of population counting. Characteristic extraction fuses a prospect characteristic, an edge characteristic and a texture characteristic. Because of the fusion characteristic, image information expression validity is increased and simultaneously characteristic correlation is generated. Besides, a problem that training data is imbalance may exist in actual population counting application. The characteristic fusion is adopted in the invention to increase an image information expression capability; the sparse constraint is used, which is good for selecting a distinguishable characteristic; a cost-sensitive learning method is introduced to reduce an adverse effect of the imbalance training data on the model so that a new high-efficiency robust cost-sensitive sparse linear regression model image population counting method is provided. An image to be tested is input and robust population counting can be rapidly realized through using a model parameter acquired through the training.

Description

A kind of robust people counting method returned based on cost-sensitive sparse linear

Technical field

The present invention relates to a kind of robust people counting method returned based on cost-sensitive sparse linear, belong to technical field of video image processing。

Background technology

In recent years, crowd's counting based on video also exists many potential using values in actual applications, including public safety (accident detection, the management that crowd flows to, the control of crowd density) market profit (determining the rent rate of retail shop according to flow of the people) urban planning (reasonably carrying out roading according to flow of the people) etc.。Therefore crowd's counting is an important research topic in technical field of video image processing。

Essence based on crowd's counting technology of video is the number of the people of appearance in automatic marking video image。But, in practical application, scene video be usually present serious block, view transformation and the problem such as video resolution is low so that the crowd's counting technology performance based on directly detection and tracking pedestrians can not practical requirement。Therefore, the people counting method studying new robust has important theory significance and actual application value。At present, the crowd based on video image counts the method that main stream approach is employing regression model, and the method based on regression model is substantially that study one is from characteristics of image to the mapping relations of number label。

The existing pedestrian counting method based on regression model includes two main modules, i.e. characteristic extracting module and model training module。For characteristic extracting module, the method for main flow is to obtain better performance by fusion prospect, edge and textural characteristics。There is problems of, direct Feature Fusion may produce feature correlation, reduces the ga s safety degree of feature。For model training module, existing multiple regression models are used for solving people counting problem, and wherein the people counting based on ridge regression (RidgeRegression, RR) model has been achieved with good performance。Studies have found that the pedestrian counting method robustness based on ridge regression model is not enough, poor performance can be produced under the conditions such as training data imbalance。In order to obtain the people counting technology of robust, the present invention proposes a kind of new robust people counting method based on cost-sensitive sparse linear regression model。

Cost sensitive learning is the problem solving data model error of fitting and model error by introducing different costs to estimate, and what usual cost estimated description is divide the cost needed for sample by mistake。The research of cost sensitive learning has been successfully applied in unbalanced image classification problem。Document (N.H.VoandY.Won, " Classificationofunbalancedmedicaldatawithweightedregular izedleastsquares; " inFrontiersintheConvergenceofBioscienceandInformationTec hnologies, 2007, pp.347-352.) extend the method for least square of regularization and punish that different samples is with different weights。Document (Y.Sun, M.S.Kamel, A.K.Wong, andY.Wang, " Cost-sensitiveboostingforclassificationofimbalanceddata, " PatternRecognition, vol.40, pp.3358-3378,2007.) proposing the boosting algorithm of a kind of cost-sensitive, this algorithm is by introducing cost item in the learning framework of AdaBoost。These two work significantly reduce the adverse effect of the model error brought due to uneven training sample, improve the robustness of disaggregated model。

It is subject to the inspiration of above-mentioned work, innovatively proposes a kind of robust people counting method returned based on cost-sensitive sparse linear herein。In order to reach this target, we have proposed the framework of the crowd of two-layer counting。At ground floor framework, it is contemplated that sparse constraint can select ga s safety degree feature, and the pedestrian counting method based on linear regression model (LRM) can obtain good performance, and the present invention proposes to utilize a kind of sparse linear regression model of training data study。On this basis, the label estimation difference of each training data can be obtained。At second layer framework, it is contemplated that the complex scene problem of practical application, training data distribution is likely to imbalance, and the label estimation difference obtained based on ground floor framework can be distributed difference, and big estimation difference causes that crowd's counting properties declines。Therefore, cost sensitive learning method is adopted to eliminate owing to model error brings crowd's counting properties decline problem。The robust performance of method that the present invention proposes has obtained substantial amounts of experimental verification, wherein, adopts UCSD and the Mall database authentication effectiveness of the inventive method of International Publication。

Summary of the invention

The present invention, towards the intelligent monitoring technology based on video, has invented a kind of robust people counting method returned based on cost-sensitive sparse linear, has improve crowd's counting accuracy。The method comprises the steps

A) adopt mixed Gaussian background modeling that crowd's image set is extracted foreground area；

B) visual angle rectification is done in the region that step (a) is extracted, and extracts foreground features, edge feature and textural characteristics respectively；

C) it is a characteristic vector x by three kinds of Feature Fusion in step (b)_i, and this characteristic vector is normalized operation；

D) i-th the image feature vector x extracted by step (c) is utilized_iThe label y corresponding with this picture_i, train a sparse linear regression model, and calculate the label estimation difference of every image；

E) according to the label estimation difference obtained by step (d), for the weight factor that each sample design is corresponding；

F) weight factor and one cost-sensitive sparse linear regression model of training sample re-training that step (e) obtains are utilized；

G) for the image to be tested of an input, step (a)-(c) is utilized to extract test image feature vector crowd's counting of the model parameter estimation test image obtained according to step (f)。

The beneficial effects of the present invention is: method of the present invention not only has good robustness, and can solve the problem that the crowd's enumeration problem under the distribution of training data imbalance。Our experiments show that, the method that the present invention proposes three evaluation index-mean absolute errors (MAE), Averaged Square Error of Multivariate (MSE) and all can obtain relatively low improper value under average offset error (MDE)。Test result indicate that, even if in the unbalanced situation of Crowds Distribute, the result of crowd's counting of the method that invention proposes still can reach the result close with artificial counting result。Simultaneously the algorithm complex of the present invention is low, and hardware requirement level is low, it is easy to real time execution is in the limited environment of operational capability。

Accompanying drawing explanation

Fig. 1. based on the robust people counting method theory diagram that cost-sensitive sparse linear returns

Fig. 2. SLR model and CS-SLR model performance contrast schematic diagram when training data imbalance

Fig. 3. crowd counts actual value and the inventive method crowd counts estimated value comparison diagram

Fig. 4. crowd's counting algorithm Performance comparision

Detailed description of the invention

Below in conjunction with the drawings and specific embodiments, the present invention is described in further detail。

1. the extraction of characteristic vector

Characteristic vector pickup is the important module during crowd counts。In order to obtain the people counting method of a robust, the present invention adopts the method that multi-domain characteristics merges, and namely combines following three kinds of features and forms a characteristic vector:

Foreground features: one of feature that people counting method that foreground features is is the most frequently used, is characterized by that the method modeled by Gaussian Background is obtained。Saying physical significance, this feature description is shape and the size of crowd。According to the prospect obtained, we extract the features such as area, girth, perimeter edge direction and block number and represent foreground features。

Edge feature: edge feature can describe crowd's number in the foreground effectively。Rim detection is adopted to come the edge in extraction prospect, it is thus achieved that edge length, the direction at edge and Minkowski dimensional characteristics。

Textural characteristics: textural characteristics is based on gray level co-occurrence matrixes (Gray-LevelCo-occurrenceMatrix, GLCM) and obtains, and textural characteristics contains the information of important crowd's number。Extracting the homogeneity of GLCM, the feature such as energy and comentropy is estimated respectively as slickness, concordance and Texture complication。

All training images that the present invention adopts are through visual angle normalization and change into gray level image。The present invention extract characteristic vector through centralization and standardization。

2. sparse linear regression model

The present invention adopts ridge regression (RidgeRegression, RR) model as basic linear regression model (LRM)。Consider that multiple domain fusion feature may produce feature correlation, therefore it is proposed that based on the linear regression model (LRM) of sparse constraint。For more complete description, introduce ridge regression (RR) first below。

Assume to provide n training image sample, x_iRepresent the m dimensional feature vector of the i-th image pattern extracted, y_iFor image tag, the i.e. number of the corresponding crowd of labelling。RR model can be obtained by solving following optimization problem:

$\underset{\mathrm{\β}}{\arg \min }\underset{i=1}{\overset{n}{\Σ}}\left|\right|y_i-x_i\mathrm{\β}|{|}_2^2+\mathrm{\λ}|\left|\mathrm{\β}\right|{|}_2^2---\left(1\right)$

Wherein, λ is regular parameter, and β is RR model parameter vector。Research shows that RR solution to model is stable。

Consider that present invention employs fusion feature represents training image, it is proposed that adopt sparse constraint method, obtain sparse linear recurrence (SparseLinearRegression, SLR) model by solving following optimization problem:

$\underset{\mathrm{\β}}{\arg \min }\underset{i=1}{\overset{n}{\Σ}}\left|\right|y_i-x_i{\mathrm{\β}}_1|{|}_2^2+{\mathrm{\λ}}_1|\left|{\mathrm{\β}}_1\right|{|}_1+{\mathrm{\λ}}_2\left|\right|{\mathrm{\β}}_1|{|}_2^2---\left(2\right)$

Wherein, λ₁And λ₂Respectively regular terms parameter, β₁It it is the parameter vector of sparse linear regression model。Obviously, contrast equation (1) and (2), β₁Openness higher than β。Sparse constraint makes sparse linear regression model can have good noise robustness。

There is many algorithms in solution formula (2), the present invention adopts elastic net method to solve。Formula (2) is converted into following form

$\underset{{\mathrm{\β}}^{*}}{\arg \min }\left|\right|y^{*}-X^{*}{\mathrm{\β}}^{*}|{|}_2^2+\mathrm{\γ}|\left|{\mathrm{\β}}^{*}\right|{|}_1---\left(3\right)$

The parameter vector of sparse linear regression model can be calculated by below equation

$\begin{array}{c}{\mathrm{\β}}_{SLR}=(1+{\mathrm{\λ}}_2)\mathrm{\β}=(1+{\mathrm{\λ}}_2)\frac{1}{\sqrt{1+{\mathrm{\λ}}_2}}{\mathrm{\β}}^{*}\\ =\sqrt{1+{\mathrm{\λ}}_2}{\mathrm{\β}}^{*}\end{array}---\left(4\right)$

The detail of elastic net method solution formula (2) is referred to article (H.ZouandT.Hastie, " Regularizationandvariableselectionviatheelasticnet; " JournaloftheRoyalStatisticalSociety:SeriesB (StatisticalMethodology), vol.67, pp.301-320,2005.)

3. cost-sensitive sparse linear regression model

Research shows that the robustness of SLR model is not enough, because the cost function of SLR model is to minimize global error, therefore, and the sample label estimation difference that the SLR model of training can be bigger to little classification this generation of test specimens under the conditions such as training data imbalance。RR linear regression model (LRM) there is also similar problem equally。Crowd in order to obtain robust counts, the thought of uneven for cost sensitive learning process classification problem has been incorporated into the framework of recurrence learning by the present invention, proposing a kind of new cost-sensitive sparse linear regression model (Cost-SensitiveSLR, CS-SLR), mathematical description is as follows:

$\underset{\mathrm{\β}}{\arg \min }\underset{i=1}{\overset{n}{\Σ}}{c}_i\left|\right|y_i-x_i\mathrm{\β}|{|}_2^2+{\mathrm{\λ}}_1|\left|\mathrm{\β}\right|{|}_1+{\mathrm{\λ}}_2\left|\right|\mathrm{\β}|{|}_2^2---\left(4\right)$

Wherein, λ₁And λ₂Respectively regular terms parameter, c_iFor the associated weight factor corresponding to i-th training sample。β is the parameter vector of cost-sensitive sparse linear regression model。

The basic thought of the present invention is to retrain the negative effect that little classification estimation difference produces。For this, the present invention adopts following exponential function to determine the weight factor c of each training image sample_i, and follow the principle of the little classification higher weight factor of acquisition, it may be assumed that

c_i=exp (| y_i-x_iβ|)/Z(5)

Wherein,Being the SLR model parameter vector estimated by formula (2), Z is normalization factor。

It should be noted that and method assumes that big estimation difference is corresponding to big c_i, this method is comparatively sensitive to exceptional value, and in order to reduce the negative effect that exceptional value produces, the present invention adopts preprocess method to remove exceptional value。Document (J.N.Miller, " Basicstatisticalmethodsforanalyticalchemistry.Part2.Cali brationandregressionmethods.Areview; " Analyst, vol.116, pp.3-14,1991.) point out that the confidence interval of data is all set to 95%, the data outside confidence interval are typically considered exceptional value, will disallowable training sample set, the present invention adopts same confidence interval。

CS-SLR model solve employing elastic net method, it is possible to be converted to following optimization problem

$\underset{\mathrm{\β}}{\arg \min }\underset{i=1}{\overset{n}{\Σ}}\left|\right|\sqrt{c_i}{y}_i-\sqrt{c_i}{x}_i\mathrm{\β}|{|}_2^2+{\mathrm{\λ}}_1|\left|\mathrm{\β}\right|{|}_1+{\mathrm{\λ}}_2\left|\right|\mathrm{\β}|{|}_2^2---\left(6\right)$

For the image pattern to be tested of an input, extracting the features training of this image pattern according to enforcement step 1, then utilize the model parameter vector β that formula (6) obtains, it is achieved the label of this test sample is estimated, namely crowd's number is estimated。

Claims (5)

1. the robust people counting method returned based on cost-sensitive sparse linear, comprises the steps:

A () adopts mixed Gaussian background modeling that crowd's image set is extracted foreground area；

B region that step (a) is extracted by () is done visual angle and is corrected, and extracts foreground features, edge feature and textural characteristics respectively；

C three kinds of Feature Fusion in step (b) are a characteristic vector x by ()_i, and this characteristic vector is normalized operation；

D () utilizes i-th the image feature vector x extracted by step (c)_iThe label y corresponding with this picture_i, train a sparse linear regression model, and calculate the label estimation difference of every image；

E label estimation difference that () basis is obtained by step (d), for the weight factor that each sample design is corresponding；

F () utilizes the weight factor and one cost-sensitive sparse linear regression model of training sample re-training that step (e) obtains；

G (), for the image to be tested of an input, utilizes step (a)-(c) to extract test image feature vector crowd's counting of the model parameter estimation test image obtained according to step (f)。

2. a kind of robust people counting method returned based on cost-sensitive sparse linear according to claim 1, it is characterised in that adopt multi-domain characteristics to merge in described step (b)；Wherein, the foreground features of employing is area respectively, girth, perimeter edge direction and block number；Edge feature is edge length respectively, edge direction and Minkowski dimension；Textural characteristics is based on the homogeneity of gray level co-occurrence matrixes, energy and entropy respectively。

3. a kind of robust people counting method returned based on cost-sensitive sparse linear according to claim 1, it is characterised in that the sparse linear regression model introducing sparse constraint acquisition crowd's counting in described step (d) is expressed as:

$\underset{\mathrm{\β}}{\arg \min }\underset{i=1}{\overset{n}{\Σ}}\left|\right|y_i-x_i{\mathrm{\β}}_1|{|}_2^2+{\mathrm{\λ}}_1|\left|{\mathrm{\β}}_1\right|{|}_1+{\mathrm{\λ}}_2\left|\right|{\mathrm{\β}}_1|{|}_2^2$

Wherein β is the parameter vector of linear regression model (LRM), λ₁And λ₂Respectively sparse regular parameter and stability parameter。

4. a kind of robust people counting method returned based on cost-sensitive sparse linear according to claim 1, it is characterized in that, label estimation difference in described step (e) is relevant to training data categorical measure, in order to reduce the adverse effect that crowd is counted by these errors, the present invention is to the different weight factor c of each training sample design_i

c_i=exp (| y_i-x_iβ|)/Z

Wherein,Being the parameter vector of the sparse linear regression model estimated, Z is normalization factor。

5. a kind of robust people counting method returned based on cost-sensitive sparse linear according to claim 1, it is characterised in that described step (f) trains new robust crowd to count cost-sensitive sparse linear regression model:

$\underset{\mathrm{\β}}{\arg \min }\underset{i=1}{\overset{n}{\Σ}}{c}_i\left|\right|y_i-x_i\mathrm{\β}|{|}_2^2+{\mathrm{\λ}}_1|\left|\mathrm{\β}\right|{|}_1+{\mathrm{\λ}}_2\left|\right|\mathrm{\β}|{|}_2^2$

Wherein, c_iThe weight factor that each training sample calculated for step (e) is relevant。