CN101430757A - Method for acquiring action classification by combining with spacing restriction information - Google Patents

Abstract

The invention discloses a method for acquiring an action category by combining space constraint information, relates to the automatic monitoring field, and solves the problem of long training time and low classification precision of the existing methods for acquiring the action categories. The method comprises the following steps: reading video, tracking a target in a target profile section by snake and a particle filter, and accurately framing the target section with a rectangular frame in each frame; acquiring a target curve and a fitting function according to the width and height of the target section of each frame, and acquiring a prior probability of a current action a which is classified as a k category action by classifying support vector machines of an extracted characteristic value, dividing the video into m sections, and acquiring the coverage probabilities of all the frames in the video that a section I is trained by a k category action training set, and the probability sum of all the categories of actions in the section i; and acquiring the category number to which the current action a belongs according to the proportion of the times that the current action a covers the section i in all the times that all the video sections are covered.

Description

A kind of method of acquiring action classification by combining with spacing restriction information

Technical field

The present invention relates to automatically-monitored field, be specifically related to obtain the method for the classification of motion in conjunction with space constraint information.

Background technology

Abnormal behaviour detection for member in the family is that the research field of a hot topic, the especially nurse to old man, child and handicapped disabled person have great significance in recent years.But present intelligent family monitoring system all is based on sensor network and Wireless Telecom Equipment mostly, and cost and cost are higher, and is not suitable for general family and uses, so, based on the wired home monitoring technique of computer vision in recent years by people's extensive concern.But the method for this class technology is very limited at present, and great majority are based on the detection of ad hoc rules and specific exceptions behavior, and the generalization ability difference causes it to be not easy to promote.There is the long and not high shortcoming of nicety of grading of training time in the present method of obtaining the classification of motion.

Summary of the invention

In order to solve the long and not high problem of nicety of grading of the method training time of obtaining the classification of motion at present, a kind of method of acquiring action classification by combining with spacing restriction information is proposed now

The step of the method for the invention is:

Step 1, read video, utilize snake and particle filter that the target in the objective contour zone is followed the tracks of, and in each frame, all use an accurate frame of rectangle frame to live the target area;

Step 2, obtain aim curve according to the width w (t) and the height h (t) of the target area of each frame $R\left(t\right)=\frac{w\left(t\right)}{h\left(t\right)};$

Step 3, aim curve R (t) is transformed to 2 π is the Fourier leaf-size class fitting function in cycle $f\left(t\right)=a_0+\underset{f=1}{\overset{\∞}{\mathrm{\Σ}}}{a}_f\cos \mathrm{ft}+b_f\sin \mathrm{ft},$ Wherein $\begin{array}{c}{a}_n=\frac{1}{\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}f\left(x\right)\cos \mathrm{nxdx}(n=\mathrm{0,1,2},L),\\ b_n=\frac{1}{\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}f\left(x\right)\sin \mathrm{nxdx}(n=\mathrm{1,2,3}L)\end{array},$ Extract f=0 respectively, 1,2,3,4 o'clock a _fAnd f=1,2,3,4,5 o'clock b _fA proper vector that is combined into;

Step 4, by the proper vector of extracting being carried out the support vector machine classification, obtain the prior probability P (e that current action a is classified as the action of k class ^k), video is divided into m zone, obtain regional i and concentrated the Probability p (i/e that all frames covered in all videos by k class action training ^k) and the probability that in regional i, occurs of all classification actions and

Wherein k ∈ 1,2...n}, i ∈ 1,2...m};

Step 5, the ratio p (i/a) that accounts for the total degree that all video areas are capped in the training set according to the number of times of current action a overlay area i obtain the classification number of action under the current action a:

$C=\arg \underset{k}{\max }\underset{i=1}{\overset{m}{\mathrm{\Σ}}}p(i/a)p(e^k/i)$

Advantage of the present invention is: it combines the rule that specific action has space constraint (1), overcome in the past based on multiclass classification of motion poor effect in the supervisory system of vision, the shortcoming that similar movement is obscured easily, because the erratic behavior of abnormal behaviour scene and the height regularity of regular event have very big difference, so the present invention has improved the precision that abnormal behaviour detects; (2) by the embedded space constraint information, having solved needs the problem of sampling in a large number and training before the motion detection system detects in the past, have the sorter fast convergence rate, need the few advantage of sample size.

Embodiment

The step of the described method of present embodiment is:

Step 1, read video, utilize snake and particle filter that the target in the objective contour zone is followed the tracks of, and in each frame, all use an accurate frame of rectangle frame to live the target area;

Step 2, obtain aim curve according to the width w (t) and the height h (t) of the target area of each frame $R\left(t\right)=\frac{w\left(t\right)}{h\left(t\right)};$

Step 3, aim curve R (t) is transformed to 2 π is the Fourier leaf-size class fitting function in cycle $f\left(t\right)=a_0+\underset{f=1}{\overset{\∞}{\mathrm{\Σ}}}{a}_f\cos \mathrm{ft}+b_f\sin \mathrm{ft},$ Wherein $\begin{array}{c}{a}_n=\frac{1}{\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}f\left(x\right)\cos \mathrm{nxdx}(n=\mathrm{0,1,2},L),\\ b_n=\frac{1}{\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}f\left(x\right)\sin \mathrm{nxdx}(n=\mathrm{1,2,3}L)\end{array},$ Extract f=0 respectively, 1,2,3,4 o'clock a _fAnd f=1,2,3,4,5 o'clock b _fA proper vector that is combined into;

Step 4, by the proper vector of extracting being carried out the support vector machine classification, obtain the prior probability P (e that current action a is classified as the action of k class ^k), video is divided into m zone, obtain regional i and concentrated the Probability p (i/e that all frames covered in all videos by k class action training ^k) and the probability that in regional i, occurs of all classification actions and

Wherein k ∈ 1,2...n}, i ∈ 1,2...m};

Step 5, the ratio p (i/a) that accounts for the total degree that all video areas are capped in the training set according to the number of times of current action a overlay area i obtain the classification number of action under the current action a:

$C=\arg \underset{k}{\max }\underset{i=1}{\overset{m}{\mathrm{\Σ}}}p(i/a)p(e^k/i)$

P (e wherein ^k/ probability that i) action of all k classes took place on i video area in the expression training set, it be by $p(e^k/i)=\frac{p(i/e^k)P\left(e^k\right)}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}p(i/e^j)p\left(e^j\right)}$ Obtain.The method that the present invention obtains the classification of motion has merged svm classifier device and the corresponding spatial information of such action, reaches the spatial information that takes place with specific action and revises the interfacial purpose of SVM.

Claims (1)

1, a kind of method of acquiring action classification by combining with spacing restriction information is characterized in that its step is:

Step 1, read video, utilize snake and particle filter that the target in the objective contour zone is followed the tracks of, and in each frame, all use an accurate frame of rectangle frame to live the target area;

Step 2, obtain aim curve according to the width w (t) and the height h (t) of the target area of each frame $R\left(t\right)=\frac{w\left(t\right)}{h\left(t\right)};$

Step 3, aim curve R (t) is transformed to 2 π is the Fourier leaf-size class fitting function in cycle $f\left(t\right)=a_0+\underset{f=1}{\overset{\∞}{\mathrm{\Σ}}}{a}_f\cos \mathrm{ft}+b_f\sin \mathrm{ft},$ Wherein $\begin{array}{c}{a}_n=\frac{1}{\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}f\left(x\right)\cos \mathrm{nxdx}(n=\mathrm{0,1,2},L),\\ b_n=\frac{1}{\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}f\left(x\right)\sin \mathrm{nxdx}(n=1,2,3L)\end{array},$ Extract f=0 respectively, 1,2,3,4 o'clock a _fAnd f=1, the proper vector that 2,3,4,5 o'clock bf is combined into;

Step 4, by the proper vector of extracting being carried out the support vector machine classification, obtain the prior probability P (e that current action a is classified as the action of k class ^k), video is divided into m zone, obtain regional i and concentrated the Probability p (i/e that all frames covered in all videos by k class action training ^k) and the probability that in regional i, occurs of all classification actions and

Wherein k ∈ 1,2...n}, i ∈ 1,2...m};

Step 5, the ratio p (i/a) that accounts for the total degree that all video areas are capped in the training set according to the number of times of current action a overlay area i obtain the classification number of action under the current action a:

$\text{C=arg}\underset{k}{\max }\underset{i=1}{\overset{m}{\mathrm{\Σ}}}p(i/a)p(e^k/i).$