CN104408747B - Human motion detection method suitable for depth image - Google Patents

Abstract

The invention provides a human motion detection method suitable for a depth image. The method comprises the following steps: firstly, dividing the image into an upper layer and a lower layer, wherein the upper layer and the lower layer are used for creating a background model by different neighborhoods, and a reference model is added while the background model is created; secondly, adjusting the parameter of the difference threshold value of an image lower-layer algorithm and comparing pixels with the background model for pixel classification in a following video; thirdly, updating the background model by different updating modes based on the classified pixels; finally, denoising false detection points. According to the method, the identification and detection rates of a human body are remarkably increased.

Description

A kind of body movement detection method being applied to depth image

Technical field

The invention belongs to computer vision field, it is related to a kind of method of the detection of the human motion for depth image.

Background technology

In the research of human motion visual analysises, human motion detection is a crucial pre-treatment step, directly affects The follow-up effect followed the tracks of, identify, therefore human motion detection algorithm is always the study hotspot in this field.

3D sensor with Microsoft Kinect as representative, can obtain the depth image of the three-dimensional information assuming object, give people Body motion detection and analysis provide new way.Compare common color image, depth image has some obvious advantages, such as tired Shade and the lighting issues of disturbing coloured image are little on depth image impact.

ViBe (visual background extractor) algorithm is also referred to as visual background and extracts Operators Algorithm, is One of background subtraction method.It is a kind of Pixel-level video background modeling algorithm, effect be better than conventional average background model, Mixed Gauss model scheduling algorithm, has the features such as amount of calculation is little, treatment effeciency is high.But due to depth image and common color figure As the difference of characteristic, the method is applied to during depth image has problems with：1) Near Ground moving target is difficult to detect, tool Embody and like the foot disappearance being connected with ground in image.2) there is the people of motion, ViBe algorithm meeting in the background image when modeling People is initialized as background, after people's motion, this region can be judged to this phenomenon of prospect " ghost ", on the other hand when people transports again Move during " ghost " region due to during overlapping overlapping cannot detect this phenomenon of prospect be referred to as " shadow ".3) due to The inherent shortcoming of Kinect sensor leads to static pixel to be mistaken for prospect.It is primarily due to the accuracy error of sensor：Long distance From pixel also have certain change in neighbouring video sequence and even lose, object edge is unstable.

Content of the invention

In order to tackle problem mentioned above, the present invention propose on the basis of classical ViBe algorithm a kind of for depth image Body movement detection method.

In order to realize object above, the inventive method mainly includes the following steps that：

Step (1). set up background model；

Step (2). pixel classifications；

Step (3). background model updates；

Step (4). flase drop point denoising Processing.

The present invention has the advantages that：

1st, a kind of adaptive image layered process and the modeling pattern of different neighbo r pattern is proposed in background modeling, And increased the reference model M of removal " ghost " phenomenon in background model_R(x).

2nd, increased foreground point testing sequence during pixel classifications, " ghost is eliminated by the comparison of current pixel and reference model Shadow ".

3rd, increased the background model more New Policy based on foreground point in terms of model modification, solve " shadow " phenomenon and ask Topic, keeps the accuracy of background model.

4th, flase drop point denoising Processing has been carried out using threshold method to classification results.

5th, complete in depth image can extract movement human, create for follow-up research such as gait analysises etc. Premise, has broad application prospects in human motion analysis field.

Brief description

Fig. 1 DViBe algorithm flow chart；

Fig. 2 a 24- neighborhood sample graph；

Fig. 2 b 14- neighborhood sample graph；

Fig. 3 difference threshold R_bImpact to PCC.

Specific embodiment

The body movement detection method of the depth image of DViBe algorithm of the present invention is described below in conjunction with the accompanying drawings.

Fig. 1 is DViBe algorithm flow chart, and alms giver will include following step in fact：

(1) divide the image into upper and lower two-layer, upper and lower two-layer sets up background model using different neighborhoods.Set up background model While increased a reference model M_R(x).

(2) difference threshold R of adjustment image lower floor algorithm_bParameter.

(3) in ensuing video, each pixel is compared with background model and carry out pixel classifications.

(4) different update modes are adopted to update background model based on sorted pixel.

(5) flase drop point denoising Processing.

One by one each step is described in detail below.

Step one, Background Modeling

(1) the adaptive image layered technology based on depth image

The essence of the demixing technology of image is by image layered ground area and non-ground area two parts.In order to adapt to Kinect The motor angle change of sensor and the change of position.Using depth image ground pixel depth value longitudinally upward direction increase The maximum distance D of characteristic and the effective sighting distance of sensor is layered to image, and detailed process is as follows：

1) string randomly selecting image starts to travel through vertically upward pixel from nethermost pixel.

2) vertical coordinate of the pixel in D ± 5 scope for first pixel value of record is y, does not have after such as having traveled through this row pixel Find the pixel in D ± 5 scope for the pixel value, then acquiescence y is the maximum of vertical coordinate.

3) process of repetition 1 and 2 m time, can obtain m ordinate value { y₁,y₂,…,y_m-1,y_m}.

4) consider object stop and sensor error situations such as the value of y that can cause excessive or too small to dividing Impact, this m ordinate value is ranked up by size, chooses k middle value, calculates their average it is possible to obtain The marginal vertical coordinate of image:

(2) pixel background model and reference model are set up

DViBe algorithm uses the mode of single-frame imagess modeling, is therefore image in the first two field picture and sets up initially Change background model.Note v (x) be image in be located at x at pixel given color space value, can by background model each Pixel x is modeled as set M (x).Value v of the n pixel randomly selecting from the neighborhood of pixel x is comprised in M (x)_i, i= 1 ..., n, referred to as background sample value, that is,

M (x)={ v₁,v₂,v₃,…,v_n-1,v_n} (2)

M_R(x)=v (x) (3)

Wherein v_iIt is the background sample value for i for the index, subscript n.For number of samples, v (x) is the pixel value of pixel x.For What upper layer images were chosen is 24 neighborhoods as Fig. 2 a, is 14 neighborhoods as Fig. 2 b for lower image selection.

Step 2, pixel classifications

(1) classification ultimate principle

By corresponding with background model for pixel value v (x) model M (x) is compared to pixel x in present image Classification.Sample value v in note v (x) and M (x)_iIn the distance of given color space it is:

dis[v(x),v_i]=| | v (x)-v_i|| (4)

To given difference threshold R.Statistics dis [v (x), v_i] number of ＜ R being represented with C, as C ＜ C_minWhen then x be Foreground point, on the contrary it is background dot, C_minIt is pixel classifications match parameter.

(2) parameter adjustment

Parameter in image layered rear pixel classifications is also required to suitably adjust, and the top section and underclad portion of image is adopted With different distance thresholds.For upper layer images classification, distance threshold R_tKeep not being changed into 20.Ground image surface due to lower image Plain depth value is close, wants effective detection at ground and goes out moving target, except change sample choose domain pattern in addition in addition it is also necessary to Adjust threshold value R_bSize.R_bLess moving target is more easily detected, but R_bToo small a large amount of background dots can be caused to be mistaken for before Sight spot.Different R_bThe impact to correct recognition rata (PCC) for the value as shown in figure 3, R can be obtained by figure_bBigger discrimination is higher, but After being greater than 6, PCC tends towards stability, and considers overall discrimination and the discrimination of foot may be selected R_b=6.

(3) foreground point verification

There is the people of motion, ViBe algorithm can be by the background that is initialized as of human factor error, therefore in the background image when modeling Generate the problem that：After people is mobile, the value of former area pixel point there occurs huge change, and has the characteristics that pixel value becomes big. These points can be judged to foreground point by ViBe algorithm, and always remains as the state of prospect.We will be this not corresponding actual The prospect point set of Moving Objects is referred to as " ghost ".Become big characteristic using pixel value to pass through in current pixel and background model Respective pixel compares, and can go ghost phenomenon.Verification rule is as follows：

Wherein v (x, t₀) it is M_RX the value of pixel x in (), v (x, t) is the value of pixel at corresponding x in present image.When away from From when exceeding Kinect sensor detection range, pixel it is possible that the situation of loss is pixel value is 0, can lead by this phenomenon Some foreground points are caused to be mistaken for background dot.Add qualificationss v (x, t₀) ＞ 0 can prevent above-mentioned reason and by normal prospect Point is mistaken for background dot.

Step 3, background model

(1) former algorithm model updates ultimate principle

The purpose that background model updates is to ensure that As time goes on background model still can keep accuracy.Work as picture When plain x is classified as background dot, just trigger the renewal process of this pixel background model M (x).Side initially with random sub-sample Method chooses whether to update M (x), for the model choosing renewal, then chooses a sample value current pixel at random from M (x) Value v (x) replace so that in background model sample value life cycle exponentially monotonic decay.In order to keep Space Consistency, Renewal process is also updated to the background model of the neighborhood territory pixel of x at random using same method.

Shadow phenomenon is because when people is re-moved to ghost region due in overlapping pixel value and its background model Sample value is close, and overlapping region can be detected as background by algorithm.In former algorithm, the more New Policy of the background model based on background dot exists Also can solve shadow in subsequent frame and but ghost phenomenon process is relatively slow.

(2) it is based on foreground point background model to update and reference model renewal

Background model based on foreground point more New Policy specific strategy is as follows：

1) in statistical picture, all pixels are continuously judged to prospect number of times F.Then again count when being detected as background dot.

2) frame number threshold value F is set_min, as F ＞ F_minShi Ze is considered that background dot is judged to foreground point by algorithm by mistake.

3) this point is revised as background dot, F is counted again, update reference model M_RX (), sets up new background model, Modeling sample is chosen using adaptive layered neighbo r pattern above.

Step 4, flase drop point denoising Processing

The feature of Kinect sensor be error increase with the increase of distance and distance less than during certain value also no Method detects.Noise occurs in that depth value is larger and object edge mostly.So substantial amounts of mistake can be removed with threshold method Cautious.Then think that this point is background when the depth value of the pixel of the prospect that is detected as exceeds set point, that is,

Wherein T and t represents depth value, and 255 represent prospect, and 0 represents background.T and t can be according to the effective sighting distance of sensor Depth value determines, if it is known that region residing for human body can determine the value of T and t according to human region, so can reduce more Flase drop point.

Experiment shows, the human motion detection that the Vibe algorithm after improvement is directed to depth image has certain feasibility, The discrimination of human body and verification and measurement ratio are all significantly improved.

Claims (3)

1. a kind of body movement detection method being applied to depth image is it is characterised in that the method comprises the following steps：

Step one, Background Modeling

(1) adaptive image layered based on depth image

Long distance using the depth image ground pixel depth value characteristic that direction increases longitudinally upward and the effective sighting distance of sensor From D, image is layered, detailed process is as follows：

1) string randomly selecting image starts to travel through vertically upward pixel from nethermost pixel；

2) vertical coordinate of the pixel in D ± 5 scope for first pixel value of record is y, does not find after such as having traveled through this row pixel Pixel value is the maximum of vertical coordinate in the pixel of D ± 5 scope, then acquiescence y；

3) repeat 1) and 2) process m time, obtain m ordinate value { y₁,y₂,…,y_m-1,y_m}；

4) this m ordinate value is ranked up by size, chooses k middle value, calculate their average, obtain image Marginal vertical coordinate：

$\stackrel{\‾}{y}=\frac{y_1+y_2+...+y_k}{k};$

(2) pixel background model and reference model are set up

Note v (x) is the depth value being located at pixel x in depth image, and each pixel x in background model is modeled as a set M(x)；Value v of the n pixel randomly selecting from the neighborhood of pixel x is comprised in M (x)_i, i=1 ..., n, referred to as background sample Value, that is,

M (x)={ v₁,v₂,v₃,…,v_n-1,v_n}；

Wherein v_iIt is the background sample value for i for the index, subscript n is number of samples；It is 24 neighborhoods for upper layer images selection, right 14 neighborhoods being chosen in lower image；If a reference model M_RX () preserves the depth value of pixel x, for foreground point school Test, that is,

M_R(x)=v (x)；

Step 2, pixel classifications

By corresponding with background model for pixel value v (x) model M (x) is compared to pixel x in present image is divided Class；Sample value v in note v (x) and M (x)_iIn the distance of given color space it is:

dis[v(x),v_i]=| | v (x)-v_i||；

To given difference threshold R；Statistics dis [v (x), v_i] number of ＜ R being represented with C, as C ＜ C_minWhen then x be prospect Point, on the contrary it is background dot, C_minIt is pixel classifications match parameter；

Step 3, background model updates

The purpose that background model updates is to ensure that As time goes on background model still can keep accuracy；When pixel x When being classified as background dot, just trigger the renewal process of this pixel background model M (x)；Method initially with random sub-sample Choose whether to update M (x), for the model choosing renewal, then choose a sample value current pixel value v at random from M (x) (x) replace so that in background model sample value life cycle exponentially monotonic decay；In order to keep Space Consistency, more New process is also updated to the background model of the neighborhood territory pixel of x at random using same method；

Step 4, flase drop point denoising Processing

Remove substantial amounts of flase drop point with threshold method；Then think when the depth value of the pixel of the prospect that is detected as exceeds set point This point is background, that is,

$v\left(x\right)=\left\{\begin{array}{cc}0& v\left(x\right)>T\left|\right|v\left(x\right)<t\\ 255& else\end{array}\right.;$

Wherein T and t represents depth value, and 255 represent prospect, and 0 represents background.

2. a kind of body movement detection method being applied to depth image according to claim 1 it is characterised in that：Image After layering, in pixel classifications, different distance thresholds are adopted to the top section of image and underclad portion；Upper layer images are divided Class, distance threshold R_tKeep not being changed into 20；Because the ground pixel depth value of lower image is close, want effectively to examine at ground Measure moving target, in addition it is also necessary to adjustable range threshold value R in addition to neighbo r pattern chosen by change sample_bFor 6.

3. a kind of body movement detection method being applied to depth image according to claim 1 it is characterised in that：When building There is the people of motion in background image during mould, become big characteristic using pixel value and pass through current pixel picture corresponding with background model Element more then removes ghost phenomenon；Verification rule is as follows：

$\left\{\begin{array}{c}v(x,t_0)-v(x,t)>0\\ v(x,t_0)>0\end{array}\right.;$

Wherein v (x, t₀) it is M_RX the value of pixel x in (), v (x, t) is the value of pixel at corresponding x in present image.