CN101996401B - Target analysis method and apparatus based on intensity image and depth image - Google Patents

Abstract

The invention provides a kind of target analysis method and apparatus based on intensity image and depth image. Described equipment comprises: main element detecting unit detects the main element of target from the intensity image of target; Foreground segmentation unit, the intensity level calculating strength threshold value in the main element based on detecting, is used this intensity threshold that intensity image is converted to binary map, and uses described binary map to carry out mask to the depth image of target, to obtain mask depth image; Movable part detecting unit detects the movable part of target from mask depth image.

Description

Target analysis method and apparatus based on intensity image and depth image

Technical field

The present invention relates to a kind of video object analytical method and equipment. More particularly, the present invention relates to onePlant the target analysis method and apparatus based on intensity image and depth image, by the method and equipment, canFrom video flowing, accurately detect all parts of interested video object, to obtain the action of targetInformation.

Background technology

Along with the extensive use of 3D animation, game, human-computer interaction interface etc., at present in computer visionAnd area of pattern recognition, just video object analytical technology is carried out to broad research. In order to carry out animation simulation,First need to from video, detect all parts of target, to obtain the action message of target

In order to detect the parts of target, conventionally use method (bottom-upmethod) from bottom to top andTop-down method (up-bottommethod). In a kind of front method, first utilize target componentFeature detect target component, as utilize the detection of skin color, for bar shaped detection of four limbs etc.,Then according to its verification model, the candidate detecting is integrated. It is right that this method depends on very muchThe exploitation of the characteristic feature of target component and the inhibition to ambient noise. In a kind of rear method, to orderTarget configuration space carries out repetitiousness search, proposes hypothesis and each hypothesis is evaluated. This method is wantedAsk the search volume must be very little, to improve computational speed.

In video object analytic process, conventionally need to from original image, be partitioned into prospect, remove mixedRandom background area, so that accurately detect target. Traditional foreground segmentation method is mainly for sourceIn the color image of CCD camera. The background of color image is generally very chaotic, therefore based on color imageThe very large and inaccuracy of foreground segmentation amount of calculation. Although also proposed based on before depth image at presentScape dividing method, but also inaccuracy of the method.

Once foreground segmentation inaccuracy, will make follow-up target analysis become very difficult, and orderMark analysis result can be subject to the interference of the ambient noise with similar features. In addition, in chaotic background oftenCan present abundant edge feature, these edge features become noise, affect target component detection or falseIf the edge feature analysis result in evaluating.

Therefore, thus need a kind ofly can cut apart more accurately the target that prospect improves target analysis performance and divideAnalysis method and equipment.

Summary of the invention

Thereby one object of the present invention is to provide one can cut apart more accurately prospect raising target dividesAnalyse the target analysis method and apparatus of performance.

Another object of the present invention is to provide a kind of target that can realize skeleton line checking accurately to divideAnalysis method and equipment.

To achieve these goals, equipment of the present invention and method are no longer confined to use color image, andIntensity image and the depth image adopting from for example TOF camera.

For realizing object of the present invention, provide a kind of target of depth image and intensity image of based targetAnalytical equipment, this equipment comprises: main element detecting unit detects target from the intensity image of targetMain element; Foreground segmentation unit, the intensity level calculating strength threshold in the main element based on detectingValue, used this intensity threshold that intensity image is converted to binary map, and use described binary map to targetDepth image carries out mask, to obtain mask depth image; Movable part detecting unit, from the mask degree of depthIn image, detect the movable part of target.

According to the present invention, the learning method training by statistics of main element detecting unit, with from intensity imageDetect the main element of target. In addition, foreground segmentation unit is to the intensity level in the main element detectingCarry out statistical average, and using income value as intensity threshold.

According to the present invention, movable part detecting unit comprises: skeleton line detecting unit, and from mask depth mapIn picture, detect many skeleton lines; Skeleton line authentication unit, verifies described many skeleton lines, to selectGo out best skeleton line. Skeleton line authentication unit is by the main element that main element detecting unit is detectedAs constant, described many skeleton lines are verified. In addition, skeleton line authentication unit is based on main bodyThe main element that parts detecting unit detects is determined the shoulder point of target, by definite shoulder point and instituteOne end of stating one of many skeleton lines connects to form dummy skeleton line, and to described many skeleton lines withAnd dummy skeleton line is verified.

In addition, the main element of described target can be people's head and trunk, and movable part can be peopleFour limbs.

To achieve these goals, provide a kind of depth image of based target and the target of intensity image to divideAnalysis method, the method comprises the following steps: the main element that a) detects target from the intensity image of target;B) the intensity level calculating strength threshold value in the main element of the target based on detecting; C) use described intensityIntensity image is converted to binary map by threshold value; D) use described binary map to carry out mask to depth image, obtainObtain mask depth image; E) from mask depth image, detect the movable part of target.

Therefore, different from the foreground segmentation based on color image in conventional art, in the present invention, firstRelatively simple main element by simple and quick method based on intensity image detection target is covered to formMould. Then, from depth image, be partitioned into prospect based on this mask, and further detect the activity of targetThe skeleton line of parts. Use statistical learning method to train the parts detector obtaining to can be used for carrying out targetMain element detects, thereby roughly estimating target position and ratio. This can significantly reduce parameter and searchRope space. In addition, in skeleton line checking, used virtual component, thereby can verify more accuratelySkeleton line.

Brief description of the drawings

Fig. 1 is the block diagram illustrating according to the structure of target analysis equipment of the present invention;

Fig. 2 illustrates the example of intensity image and depth image;

Fig. 3 illustrates the example of the main element of target;

Fig. 4 (a)-(d) illustrates respectively intensity image, binary map, depth image and mask depth imageExample;

Fig. 5 illustrates the example of the structure of movable part detecting unit;

Fig. 6 illustrates the example of the structure of skeleton line detecting unit;

Fig. 7 illustrates the example of the testing process of skeleton line detecting unit;

(a) of Fig. 8 and the example of search tree and the skeleton line the result of target (b) is shown respectively;

Fig. 9 illustrates the flow chart according to target analysis method of the present invention.

Detailed description of the invention

Below, describe in detail with reference to the accompanying drawings according to target analysis equipment of the present invention and method. The present inventionTarget analysis equipment and method can be used for from video flowing, interested video object being analyzed, withObtain the information about all parts of video object.

Conventionally,, for a video object to be analyzed, it roughly can be divided into two parts: a part isThe relatively-stationary main element such as shape, ratio, the base of for example people's head and trunk, specific deviceDeng; Another part is can activity and the relatively unfixed movable part of shape, for example people's four limbs, spyDetermine the lever arm of device etc. In the present invention, for better simply main element, can utilize through sample and uniteGrader/the detector of meter learning method training detects, and for more complicated movable part, can be throughForeground segmentation detects after removing ambient noise.

It is pointed out that for convenience, below will explain as an example of human body target example order of the present inventionMark analytical equipment and method. But, the invention is not restricted to human body target.

With reference to Fig. 1, Fig. 1 is the block diagram that the structure of target analysis equipment 100 of the present invention is shown. ThisBright target analysis equipment 100 comprises: main element detecting unit 101, foreground segmentation unit 102 andMovable part detecting unit 103. This target analysis equipment 100 from the intensity image of outside receiving target andDepth image is inputted as system. Or target analysis equipment 100 of the present invention also can comprise for catchingCatch the intensity image of target and the device of depth image, for example TOF (TimeOfFlight) camera. By forceThe albedo of degree image reflection targeted surface material, its each pixel value represents the anti-of targeted surface materialPenetrate intensity level. Distance between depth image reflection camera and target surface, its each pixel value instruction phaseMachine is to the depth value of target surface. (a) of Fig. 2 and (b) in show respectively intensity image and depth imageExample.

The main element that main element detecting unit 101 detects target for the intensity image from target is (rightIn human body target, i.e. people's head and trunk) as shown in Figure 3. Traditional detection target subject portionThe method of part is used color image, for example detection method based on background subtraction, color cluster method etc. SoAnd working strength image of the present invention is as the basis of detecting target subject parts. Intensity image can pass throughFor example TOF camera obtains.

Main element detecting unit 101 can be to utilize the statistical learning method based on sample to trainPart classification device/detector. For example, main element detecting unit 101 can useDisclosed statistics in No. US2006147108A1 and US2009097711A1 U.S. Patent applicationLearning method. Intensity image due to what input, so compared with the color image more chaotic with background, inspectionDegree of testing the speed can be accelerated, and approximately can save the time cost of half. Certainly, the invention is not restricted to this, mainBody component detecting unit 101 also can detect order by other target component detection method from intensity imageTarget main element.

The intensity image of main element detecting unit 101 scanning inputs is to determine the position of target subject partsPut, size, ratio etc. Conventionally, between the main element of target, there is more fixing geometrical-restriction relation,For example, in the time that people stands, its head and trunk have upper-lower position constraint. Based on this point, can adoptTwo kinds of strategies are searched for the main element of target. Strategy is that the institute in scanning input intensity image is fenestrateMouth is to detect a main element, and then the geometrical-restriction relation based between main element, scans thisThe adjacent image regions of main element is to detect another main element. This strategy can significantly reduce scanning window.Another strategy is that all windows in scanning input intensity image are to detect all possible parts. This planSlightly do not consider the known geometrical constraint between target subject parts. All main element candidates detectedAfterwards, if there are multiple candidates in each main element, can be according to the geometrical constraint between partsSelect best main element pair.

The main element of the target being detected by main element detecting unit 101 will divide for follow-up prospectWhen threshold calculations in cutting and movable part detect, movable part region determines.

Foreground segmentation unit 102, for remove background area (noise region) from depth image, is partitioned intoThe prospect that comprises target. As previously mentioned, traditional foreground segmentation method based on color image and based on deeplyAll inaccuracy of foreground segmentation method of degree image, bring very large difficulty to target analysis. Therefore in the present inventionIn, utilize intensity image and depth image to realize convenient and accurate foreground segmentation. Below will retouch in detailState the operation of foreground segmentation unit 102.

First,, based on the main element of main element detecting unit 101 detected targets, prospect is dividedCut unit 102 and in intensity image, calculate an intensity threshold. Preferably, can pass through in intensity image,Add up flat to be labeled as the intensity level of all pixels of main element by main element detecting unit 101All obtain this intensity threshold.

Then, foreground segmentation unit 102 by each pixel value of the intensity image of input with calculate strongDegree threshold value compares, and with by each pixel value binaryzation of intensity image, thereby intensity image is convertedFor binary map. For example, the pixel that intensity level in intensity image can be greater than to this intensity threshold is labeled as target,The pixel that intensity level is less than to this intensity threshold is labeled as background. Here can hold the binary map generating,Row medium filtering, corrosion (erosion) and (dilate) operation etc. of expanding, to remove making an uproar in binary mapSound. The example of green strength image has been shown in Fig. 4 (a), illustrated in (b) form based on threshold value twoThe example of value figure.

Afterwards, utilize this binary map to carry out mask to the depth image of input, obtain mask depth image.Binary map can be divided into depth image two parts: a part is the foreground area at target place, another portionDividing is background area. Therefore, utilize binary map to carry out mask to depth image, can delete background area.For example, can between binary map and depth image, carry out AND computing and obtain mask depth image. FigureThe example of the depth image of input is shown in 4 (c), the mask that utilizes binary map to obtain has been shown in (d)The example of depth image.

After obtaining mask depth image by foreground segmentation unit 102, movable part detecting unit 103Based on this mask depth image detect target movable part (for human body target, people's fourLimb). Due to as shown in Fig. 4 (d), in mask depth image, remove background area, chaotic limitEdge significantly reduces, and therefore movable part detecting unit 103 can detect the activity of target more accuratelyParts.

At present, movable part, the detection method of for example four limbs is broadly divided into two classes: a class is the end of based onThe method of layer feature, another kind of is method based on high-level characteristic. A kind of front method is utilized flex point, edgeThe features such as line detect the movable part of target, for example IJCV (InternationalJournalofComputerVision) " ProbabilisticMethodsforFindingPeople " and the ECCV (European in 2001ConferenceonComputerVision) " the Humanposeestimationusinglearnt in 2004Probabilisticregionsimilaritiesandpartialconfigurations " etc. A kind of rear method based onMode identification method, has adopted positive and negative sample learning grader/detector, for example MVA (MachineVisionApplications) " the Multi-viewhumanheaddetectioninstatic in 2005Images ". Movable part detecting unit 103 in the present invention can adopt said method from the mask degree of depthIn image, detect the movable part of target. Below, by the edge feature detection side with based on Hough conversionMethod is the concise and to the point describing activity parts of example detecting unit 103. It should be understood that and the invention is not restricted to this.

Can comprise two parts according to the movable part detecting unit 103 of current example: skeleton line (ridge)Detecting unit 501 and skeleton line authentication unit 502. Skeleton line detecting unit 501 detects goal activities portionThe possible skeleton line of part. Skeleton line authentication unit 502 is to the detected bone of skeleton line detecting unit 501Stringing candidate verify, to select best skeleton line. Here, skeleton line refers to each parts of targetAxis.

As shown in Figure 6, skeleton line detecting unit 501 can comprise: Depth Stratification (depthslicing) is singleUnit 601, edge detection unit 602, rectangle fitting unit 603 and skeleton line extraction unit 604.

The region at the movable part place of target is determined in Depth Stratification unit 601 in mask depth image.Here, based on by the definite main element region of main element detecting unit 101 (as institute in Fig. 7 (a)Show), the region except main element region in mask depth image can be defined as to the main part from targetThe extended movable part of part region.

In addition, likely there is such situation: in intensity image, the movable part of target and main bodyParts overlapping (people's left arm as shown in Figure 3). For this situation, Depth Stratification unit 601Calculate the mean depth value in main element region, and set it as the pixel reality of determining in main element regionBelong to the Depth Stratification threshold value of main element or movable part. This is because in depth image, asMovable part and the main element of fruit target are overlapping, movable part and main element by different deeplyIn degree plane. If the absolute value of the difference of pixel and this threshold value is greater than predetermined constant, think that this pixel belongs toIn the movable part of target, otherwise think that this pixel belongs to main element. In Fig. 7 (b), illustrate by deeplySpend the example in the definite people's in layering unit 601 four limbs region.

Edge detection unit 602 is carried out Hough conversion to depth image, obtains the edge spy of presentation graphs pictureThe Hough line of levying. Be arranged in the Hough line in the movable part region definite by Depth Stratification unit 601Represent the edge of goal activities parts. The example of the Hough line detecting has been shown in Fig. 7 (c).

Rectangle fitting unit 603 utilizes the linear orthogonal movable part with coverage goal of these Hough,As shown in Fig. 7 (d). Rectangle fitting process is roughly as follows: first, according to the Hough line detectingLength sorts to Hough line; Select the long limit of longer Hough line as rectangle, by along withThe vertical direction scan depths value trip point in depth image in long limit is determined the minor face of rectangle, thus woundBuild rectangle; If rectangle is overlapped, the principle based on deleting the little rectangle overlapping with large rectangle is deletedRedundancy rectangle.

The rectangular extraction axis that skeleton line extraction unit 604 creates from rectangle fitting unit 603, doesFor possible skeleton line, as shown in Fig. 7 (e).

Skeleton line detecting unit 501 can detect all possible movable part bone from mask depth imageStringing. These skeleton lines will be verified by skeleton line authentication unit 502 as candidate, to determine activityThe final skeleton structure of parts.

Skeleton line authentication unit 502 detects skeleton line detecting unit 501 by Model PriorSkeleton line candidate verify, to select best skeleton line, the movable part bone of final restore targetFrame. In the numerous skeleton line candidates that detect at skeleton line detecting unit 501, skeleton line authentication unit 502Which need can match with actual activity parts after checking skeleton line candidate combinations. For this reason, skeleton line checkingUnit 502 can adopt " supposing-evaluate (hypothesisgeneration-evaluation) " method. ,From all skeleton line candidates, select some skeleton lines and connect according to ad hoc fashion, then utilizing targetModel Prior and algorithm calculate the probability of this hypothesis. Select the hypothesis of maximum probability as best roadFootpath, thus detect the final skeleton line structure of target.

Skeleton line authentication unit 502 of the present invention can adopt conventional verification model at present to carry out skeleton lineChecking, pictograph structure (pictorialstructure) method (P.F.Felzenszwalb. of for example FelzenszwalbD.P.Huttenlocher, " EfficientMatchingofPictorialStructures ", CVPR2000). WithUnder, will the operation of skeleton line authentication unit 502 briefly be described as an example of above-mentioned pictograph structural approach example. ShouldUnderstand, the invention is not restricted to this.

First, skeleton line authentication unit 502 is set up the tree structure of target, for example, shown in Fig. 8 (a).This tree structure can represent the search volume of target, represents position and the annexation of all parts of target,The parameter of each node can comprise position, scale of components and the anglec of rotation etc. of parts in 3d space.

Here, can by the main element of main element detecting unit 101 detected targets together with position,Ratio and angle bisecting dispensing search tree, and set it as constant and treat. Then, skeleton line authentication unit502 by skeleton line candidate allocation to node, to produce hypothesis. Skeleton line authentication unit 502 is by belowEquation 1 calculate the observation similitude (observationsimilarity) between this hypothesis and real image:

$\mathrm{match}=w\×\frac{\mathrm{NoMatchedPt}}{(\mathrm{MatchedPt}+\mathrm{NoMatchedPt})\×\mathrm{nS}+1/\mathrm{MatchedPt}}---\left(1\right)$

Wherein, " MatchedPt " represents the non-zero pixels of the parts covering of being supposed on depth image and is somebody's turn to doThe ratio of the size of parts. " NoMatchedPt " represents zero picture of the parts covering of being supposed on depth imageThe ratio of element and the size of these parts. " nS " represents the scale factor of the rectangle of hypothesis. " w " is weight systemNumber.

Suppose that the equation (2) that the Model Matching between geometrical model can be passed through below calculates:

$d_{\mathrm{ij}}(l_i,l_j)=\underset{k=1}{\overset{3}{\mathrm{\Σ}}}{w}_{\mathrm{\θ}}^k|({\mathrm{alpha}}_j^k-{\mathrm{alpha}}_i^k)-{\mathrm{alpha}}_{\mathrm{ij}}^k|$

$+\underset{k=1}{\overset{3}{\mathrm{\Σ}}}{w}_s^k|({\log \mathrm{nS}}_j^k-\log {\mathrm{nS}}_i^k)-{\log \mathrm{nS}}_{\mathrm{ij}}^k|---\left(2\right)$

$+\underset{k=1}{\overset{3}{\mathrm{\Σ}}}{w}^k|p_{\mathrm{ij}}-p_{\mathrm{ji}}|$

Wherein, " alpha " represents the angle between adjacent component, i.e. father node and son joint in tree structureAngle between point; " alpha_ij" represent the desired angle between node i and node j; " log (nS_i ^k)”Represent the scale factor of the j parts of k dimension; " log (nS_ij ^k) " be the desired proportions factor of model; " p_ij”It is the coordinate difference of the line between parts i and parts j.

Then the posterior probability that, can suppose by equation 3 approximate calculation:

$D\left(l_i\right)=\underset{j}{\min }(d_{\mathrm{ij}}(l_i,l_j)+\mathrm{match})---\left(3\right)$

Skeleton line authentication unit 502 is found out the optimal path with maximum probability, to detect final targetSkeleton line structure. The example of skeleton line the result has been shown in Fig. 8 (b).

As mentioned above, in the present invention, skeleton line authentication unit 502 is by main element detecting unit 101The main element of detected target, together with position, ratio and angle bisecting dispensing search tree, like this canSignificantly to dwindle search volume, improve computational speed.

Because skeleton line candidate's quantity can be a lot of and corresponding method of attachment is also a lot, cause search emptyBetween very large, amount of calculation can be very large. And in main element detecting unit 101, detected orderTarget main element (for example, people's head and trunk), has determined position and the ratio of main element. RightIn specific objective type, the ratio between its each parts meets specified conditions, changes less. For example,Utilize statistical method, can be from the hand arm ratio of inferring people of people's head or trunk. Therefore, willThe main element having detected is as the known constant in search volume, and do not participate in various may combinationEnumerate, can significantly reduce search volume, reduce amount of calculation.

In addition, may there is the parts crested of target or undetected situation, people's as shown in Figure 3Right upper arm. For in this case, in skeleton line authentication unit of the present invention, also can introduce virtual portionThe concept of part, so that the parts that reduction is omitted or covered improve verifying speed and the degree of accuracy.

First, the main element detecting based on main element detecting unit 101, can determine the work of targetThe tie point of dynamic component and main element, for example people's shoulder point place. Conventionally, this tie point is positioned at orderThe step place of trough in the upright projection of target main element and horizontal projection, as shown in Figure 3.

Determining after tie point, can create a dummy skeleton line from tie point, form a hypothesis withJust evaluate checking. For example, tie point can be connected with k article of skeleton line candidate's end points, create, a virtual skeleton line candidate. This dummy skeleton line candidate and k article of skeleton line candidate form target, an arm, thus a hypothesis formed.

In this case, can ignore for the 2nd in equation 2, thereby form equation (4) below:

$d_{\mathrm{ij}}(l_i,l_j)=\underset{k=1}{\overset{3}{\mathrm{\Σ}}}{w}_{\mathrm{\θ}}^k|({\mathrm{alpha}}_j^k-{\mathrm{alpha}}_i^k)-{\mathrm{alpha}}_{\mathrm{ij}}^k|---\left(4\right)$

$+\underset{k=1}{\overset{3}{\mathrm{\Σ}}}{w}^k|p_{\mathrm{ij}}-p_{\mathrm{ji}}|$

Below, with reference to Fig. 9, target analysis method of the present invention is described.

In step S901, from the intensity image of target, detect the main element of target, as people's headAnd trunk. The intensity image of target can be obtained by TOF camera. In addition, as previously mentioned, can use based onThe parts detector of statistical learning method training detects the main element of target from intensity image.

In step S902, the main element of the target based on detecting in step S901, in intensityIn image, calculate intensity threshold. More particularly, can be by intensity image, in step S901The intensity level that is marked as all pixels of main element carries out statistical average and obtains this intensity threshold.

Next,, in step S903, use the intensity threshold calculating that intensity image is converted to two-valueFigure. At step S904, use the binary map forming at step S903 to the capable mask of the depth image of target,Thereby obtain mask depth image;

At step S905, from mask depth image, detect the movable part of target, as people's four limbs. CanTo detect the movable part of target from mask depth image by foregoing method.

Target analysis equipment of the present invention and method have more than been described. This target analysis equipment and method based onIntensity image and the depth image of target are analyzed target, can detect all parts of target, carryGet skeleton line, thus the action message of acquisition target etc. Target analysis equipment of the present invention and method are passableObtain following effect.

First, in the present invention, replace the color image of background complexity, use the intensity of target to be analyzedCarry out foreground segmentation with deep video stream, thereby can realize accurate foreground segmentation.

Secondly, the foreground segmentation result based on good, can carry out quick and accurate object edge and extract,Thereby detect all parts of target.

In addition, in the present invention, also by the statistical model detector applies based on sample in intensity image,From intensity image, successfully detect the main element of target. And the main element detecting can be appliedIn follow-up skeleton line verification system, thereby reduce the search volume of verification system, significantly reduced computingAmount.

In addition, in skeleton line verification system, also use virtual component, reducible omission or coverParts, have improved verifying speed and the degree of accuracy.

Claims (10)

1. a target analysis equipment for the depth image of based target and intensity image, this equipment comprises:

Main element detecting unit detects the main element of target from the intensity image of target;

Foreground segmentation unit, the intensity level calculating strength threshold value in the main element based on detecting, is usedIntensity image is converted to binary map by this intensity threshold, and use the depth image of described binary map to targetCarry out mask, to obtain mask depth image;

Movable part detecting unit detects the movable part of target from mask depth image,

Wherein, described equipment also comprises TOF camera, for depth image and the intensity image of captured target,

Wherein, movable part detecting unit comprises:

Skeleton line detecting unit detects many skeleton lines from mask depth image;

Skeleton line authentication unit, verifies described many skeleton lines, to select best skeleton line,

Wherein, the main element that skeleton line authentication unit detects based on main element detecting unit is determinedThe movable part of target and the tie point of main element, by definite tie point and described many skeleton lines itOne end of one connects to form dummy skeleton line, and described many skeleton lines and dummy skeleton line are enteredRow checking.

2. equipment as claimed in claim 1, is characterized in that, main element detecting unit is learned by statisticsLearning method training, to detect the main element of target from intensity image.

3. equipment as claimed in claim 1, is characterized in that, foreground segmentation unit is to the master who detectsIntensity level in body component carries out statistical average, and using income value as intensity threshold.

4. equipment as claimed in claim 1, is characterized in that, skeleton line authentication unit passes through main bodyThe main element that parts detecting unit detects, as constant, is verified described many skeleton lines.

5. equipment as claimed in claim 1, is characterized in that, the main element of target comprises people's headPortion and trunk, the movable part of target comprises people's four limbs.

6. a target analysis method for the depth image of based target and intensity image, the method comprise withLower step:

A) from the intensity image of target, detect the main element of target;

B) the intensity level calculating strength threshold value in the main element of the target based on detecting;

C) use described intensity threshold that intensity image is converted to binary map;

D) use described binary map to carry out mask to depth image, obtain mask depth image;

E) from mask depth image, detect the movable part of target,

Wherein, described method also comprises step: depth image and the intensity of utilizing TOF camera captured targetImage,

Wherein, step e) comprising:

E1) from mask depth image, detect many skeleton lines;

E2) described many skeleton lines are verified, to find out best skeleton line,

Wherein, step e2) comprising:

E21) main element based on a) detecting in step, determines movable part and the main element of targetTie point;

E22) definite tie point is connected with one end of one of described many skeleton lines, to form virtual boneStringing;

E23) described many skeleton lines and dummy skeleton line are verified.

7. method as claimed in claim 6, is characterized in that, step a) in, use by systemThe parts detector of meter learning method training detects the main element of target from intensity image.

8. method as claimed in claim 6, is characterized in that, step b) in, to what detectIntensity level in the main element of target carries out statistical average, and using income value as described intensity threshold.

9. method as claimed in claim 6, is characterized in that, at step e2) in, by will be in stepThe rapid main element detecting in a), as constant, is verified described many skeleton lines.

10. method as claimed in claim 8, is characterized in that, the main element of target comprises people'sHead and trunk, the movable part of target comprises people's four limbs.