CN102402289A - Mouse recognition method for gesture based on machine vision - Google Patents

Abstract

The invention discloses a mouse recognition method for gesture based on machine vision, comprising the following steps: (1) establishing an active shape model for gestures; (2) off-line training for gestures and getting a gesture features classifier; (3) collecting images; (4) extracting partial binary pattern characteristics for the images, searching the target gesture among the images through the gesture features classifier from step (2), then step (5); (5) positioning on the fingertips; (6) mapping of mouse. The invention is a natural and intuitive way of human-computer interaction, and has advantages of no need to carry other auxiliary equipments, finishing the mouse operation through natural hand and finger movement, small influence by light and background, and freely operation right switch when multi-users are operating the computer.

Description

A kind of gesture mouse recognition methods based on machine vision

Technical field

The present invention relates to machine vision technique and human-computer interaction technology, particularly a kind of gesture mouse recognition methods based on machine vision.

Background technology

The development of Along with computer technology, the interacting activity of people and computing machine become an important component part of human schedule life.Mouse mainly is used in the man-machine interaction of graphical interfaces; But in reciprocal process; Staff must be held a kind of mouse apparatus; On a plane, move, all there is certain limitation in this mode of operation at aspects such as the naturality of man-machine interaction and friendly, and therefore research meets the development trend that the human man-machine interaction mode of being accustomed to naturally becomes field of human-computer interaction.

Human-computer interaction technology based on gesture mainly changes the operation that judges need be carried out through identification user's gesture.It is divided into based on the receipt gloves with based on two types of machine vision.Based on the method for data glove, require the user to wear a kind of sensor of similar gloves, thereby obtain the position and the movable information of hand through computing machine, there are shortcomings such as burdensome, dumb in this method.Existing method based on machine, the method that mostly adopts the colour of skin to extract are come the dividing gesture zone, and this method receives illumination effect big, and require usually the zone of video acquisition be confined to hand region among a small circle in, this has limited the scope of activities of staff.

Summary of the invention

In order to overcome the deficiency of prior art, the object of the present invention is to provide a kind of natural, intuitively based on the gesture mouse recognition methods of machine vision.

The object of the invention is realized through following technical scheme:

A kind of gesture mouse recognition methods based on machine vision may further comprise the steps:

(1) sets up the moving shape model of gesture: at first the profile of the open gesture of the five fingers is carried out the basic gesture modeling of B batten; On the basis of basic gesture model, set up the moving shape model of the gesture of B batten then, shown in the following formula of its state space:

χ＝(x，y，θ，s，λ，θ ₁，l ₁，θ ₂，l ₂，θ ₃，l ₃，θ ₄，l ₄，θ ₅，θ ₆)

Wherein, x, y are respectively the horizontal ordinate and the ordinate of the position of gesture; θ representes the gesture anglec of rotation planar; S represents the size of gesture; θ ₁, l ₁Represent angle and the little finger length of little finger respectively around its root rotation; θ ₂, l ₂The expression third finger is around the angle and the nameless length of its root rotation respectively; θ ₂, l ₃Represent angle and the middle finger length of middle finger respectively around its root rotation; θ ₄, l ₄Represent angle and the forefinger length of forefinger respectively around its root rotation; θ ₆The expression thumb is around the angle of its root rotation; θ ₅The expression thumb is around the angle of joint, center rotation;

(2) gesture is carried out off-line training, obtain the gesture feature sorter;

(3) images acquired;

(4) extract partial binary pattern (LBP) characteristic of image, the gesture feature sorter ferret out gesture in image that obtains through step (2); When searching the target gesture, carry out step (5);

(5) finger tip location:

(5-1) adopt the moving shape model of the gesture in the step (1) that the target gesture that searches is estimated match, obtain the state space initial value χ of target gesture ₀, make the initial value p of posterior probability density ₀(χ ₀| Z ₀)=1;

(5-2) the target gesture is observed, utilized the particle filter tracking algorithm that the state of target gesture is carried out the iteration renewal, said iteration renewal process is specific as follows:

The observed quantity that defines the t-1 two field picture is Z _T-1, then the state space of target gesture is χ during the t-1 frame _T-1Posterior probability be p _T-1(χ _T-1| Z _T-1);

The optimal profile state of target gesture during according to the t-1 frame

The optimal profile state of target gesture during to the t frame

The estimation probability be: p _t(χ _t| χ _T-1); Order

The likelihood probability of the profile state of target gesture is p during the t frame _t(Z _t| χ _t);

The state space of target gesture is χ when obtaining the t frame _tPosterior probability be:

p _t(χ _t|Z _t)＝p _t(Z _t|χ _t)*∫p _t(χ _t|χ _t-1)*p _t-1(χ _t-1|Z _t-1)dχ _t-1/p _t(Z _t)；

The posterior probability density p of target gesture during according to the t frame _t(χ _t| Z _t) the optimal profile state of target gesture when estimating the t frame

(5-3) when the t frame, indicate the coordinate A of forefinger finger tip and the coordinate B of thumb finger tip on the optimal profile of target gesture;

(6) mouse mappings:

Make up right-angle triangle ABC: cross the A point and do vertical line, cross the B point and do horizontal line, intersection point is C;

If line segment AB and CB angle are α, the angle of line segment BA and CA is β, and α is less than threshold value formula α _TBe mapped as the left mouse button click action, β is less than threshold value beta _TBe mapped as and click action by mouse right button;

The mid point M of line taking section AB is mapped as the relative coordinate of mouse; If the forefinger fingertip location is A in the previous frame image ^*, the coordinate B of thumb finger tip ^*, line segment A ^*B ^*Mid point be M ^*, then directed line segment M*M is mapped as the relatively move vector of mouse on screen.

Step (2) is said carries out off-line training to gesture, obtains the gesture feature sorter, is specially:

Use 600 open one's eyes wide the mark gesture as positive sample and 1200 non-target gestures as negative sample; The opencv_traincascade program that use is increased income to be provided among the Flame Image Process class libraries OpenCV is trained.

The said gesture feature sorter ferret out gesture in image that obtains through step (2) of step (4) is specially:

All subwindows of 30 * 30 of intercepting entire image, each subwindow are eliminated non-target gesture subwindow step by step successively through 20 grades of gesture feature sorters, confirm as the target gesture through the subwindow of all 20 grades of gesture feature sorters; If in this layer search, do not find the target gesture, then with subwindow with 1.3 times of amplifications, detect through the gesture feature sorter again.

In the said ferret out gesture of step (2) process, if the previous frame image has searched the target gesture, then the hunting zone with current frame image is reduced into the zone around the target gesture region in the previous frame image, is specially:

The left margin distance is d to target gesture region if the center of the target gesture region that searches in the previous frame image is for some O, some O ₁, some O is d to the right margin distance ₂, distance is d to some O to the coboundary ₃, some O is d to the lower boundary distance ₄, the rectangular search frame is made for being the center with an O in then new region of search, and some O is 2*d to rectangular search frame left margin distance ₁, some O is 2*d to rectangular search frame right margin distance ₂, some O is 2*d to rectangular search upper frame edge circle distance ₃, some O is 2*d to rectangular search frame lower boundary distance ₄

Step (5-2) is said to be observed the target gesture, specifically observes through adaptive skin color segmentation method; Said adaptive skin color segmentation method may further comprise the steps:

(5-2-1) set up TSL space complexion model;

(5-2-2) carry out colour of skin filtering;

(5-2-3) bianry image that obtains after the colour of skin filtering is carried out dilation operation.

The said TSL of foundation of step (5-2-1) space complexion model is specially:

Is the TSL color space through following formula with the RGB color space conversion:

$\left\{\begin{array}{c}T=\frac{1}{2\mathrm{\π}}{\tan }^{-1}\left(\frac{r^{\′}}{g^{\′}}\right)+0.5\\ S=\sqrt{\frac{9}{5}(r^{\′2}+g^{\′2})}\\ L=0.299*R+0.587*G+0.114*B\\ \end{array}\right.$

Wherein $r^{\′}=(r-\frac{1}{3}),$ $g^{\′}=(g-\frac{1}{3})$

$r=\frac{R}{R+G+B},$ $g=\frac{G}{R+G+B}$

R, G, B are respectively the RGB component under the rgb color model; T, S, L are respectively the TSL component under the TSL colour model.

Step (5-2-2) is said carries out colour of skin filtering, is specially:

500 face and hand region that comprise the image of area of skin color are sampled the equal value matrix E and the covariance matrix ∑ of the two-dimentional Gaussian distribution probability distribution parameters of T and S under the estimation TSL model;

Each pixel is detected, if the C=that the T of a pixel and S component are formed (T, S) vector is lower than threshold value Threshold with the mahalanobis distance of mean vector E, thinks that then this pixel belongs to area of skin color; Said mahalanobis distance d=(C-E) ^T∑ ^-1(C-E).

Said threshold value Threshold is confirmed by following process:

According to equal value matrix E and covariance matrix ∑ estimation C=(T, S) distance of vector and mean vector E obtains initial value;

Calculate the degree of confidence confidence of each threshold value Threshold according to following formula:

$\mathrm{Confidence}=\frac{\mathrm{PosSkin}}{\mathrm{MaskArea}}*2-\frac{\mathrm{NegSkin}}{\mathrm{BgArea}}$

Wherein PosSkin is meant the quantity of the pixel of the colour of skin in the colour of skin template zone, and MaskArea is meant the total area in colour of skin template zone, and NegSkin is meant the quantity of skin pixel point in the background template zone, and BgArea is meant the area in background template zone;

The threshold value Threshold that degree of confidence is maximum is updated to the TSL complexion model parameter of current the best, and upgrades TSL space complexion model.

Compared with prior art, the present invention has the following advantages and technique effect:

1, the present invention is a kind of natural, man-machine interaction mode intuitively, and the user need not carry other utility appliance, accomplishes mouse action through the hand and the finger motion of nature;

2, particle filter tracking model under comprehensive LBP characteristic of the present invention and the ASM carries out gesture zone location and profile and follows the tracks of, and receives illumination and background influence little;

3, the present invention has overcome conventional mouse the user has been operated the restriction of degree of freedom, can make the user before camera in the 2m middle distance computing machine is controlled;

4, the present invention can be when the multi-user operation computing machine, free blocked operation power.

Description of drawings

Fig. 1 is the process flow diagram that the present invention is based on the gesture mouse recognition methods of machine vision.

Fig. 2 is the synoptic diagram of the open gesture of the five fingers.

Fig. 3 is the synoptic diagram of target gesture.

Fig. 4 is the synoptic diagram in the ferret out gesture process.

Fig. 5 is provided with synoptic diagram for the template in the complexion model.

Fig. 6 is the synoptic diagram that makes up right-angle triangle ABC in the mouse mappings process.

Fig. 7 clicks synoptic diagram for left mouse button.

Fig. 8 is for clicking synoptic diagram by mouse right button.

The synoptic diagram that Fig. 9 moves for mouse.

Embodiment

Below in conjunction with embodiment and accompanying drawing, the present invention is done to specify further, but embodiment of the present invention is not limited thereto.

Embodiment

Gesture mouse of the present invention recognition methods is based on the gesture mouse system; The gesture mouse system of present embodiment is made up of image capture module, image processing module and mouse event respond module; Image capture module comprises camera, is responsible for gathering user's image in real time and being transferred in the image processing module.

Image processing module and mouse event respond module are accomplished by computing machine, and wherein image processing module moves various image processing algorithm real-time analysis user images, are the control corresponding instruction with the movement conversion of user's hand and finger.The mouse event respond module is accepted the steering order that image processing module sends, and converts corresponding instruction into the mouse response events.

As shown in Figure 1, the gesture mouse recognition methods based on machine of present embodiment may further comprise the steps:

(1) sets up the moving shape model of gesture: at first the profile of the open gesture of the five fingers (as shown in Figure 2) is carried out the basic gesture modeling of B batten; On the basis of basic gesture model, set up the moving shape model of the gesture of B batten then, shown in the following formula of its state space:

χ＝(x，y，θ，s，λ，θ ₁，l ₁，θ ₂，l ₂，θ ₃，l ₃，θ ₄，l ₄，θ ₅，θ ₆)

Wherein, x, y are respectively the horizontal ordinate and the ordinate of the position of gesture; θ representes the gesture anglec of rotation planar; S represents the size of gesture; θ ₁, l ₁Represent angle and the little finger length of little finger respectively around its root rotation; θ ₂, l ₂The expression third finger is around the angle and the nameless length of its root rotation respectively; θ ₃, l ₃Represent angle and the middle finger length of middle finger respectively around its root rotation; θ ₄, l ₄Represent angle and the forefinger length of forefinger respectively around its root rotation; θ ₅, θ ₆The expression thumb is around the angle of its root rotation; θ ₆The expression thumb is around the angle of its root rotation; θ ₅The expression thumb is around the angle of joint, center rotation;

(2) off-line training is carried out in the staff zone, obtains the gesture feature sorter: use 600 open one's eyes wide the mark gesture as positive sample and with 1200 non-target gestures as negative sample; The opencv traincascade program that use is increased income to be provided among the Flame Image Process class libraries OpenCV is trained.

(3) images acquired;

(4) partial binary mode characteristic of extraction image is searched for target gesture as shown in Figure 3 through the gesture feature sorter that step (2) obtains in image; When searching the target gesture, carry out step (5);

The said gesture feature sorter ferret out gesture in image that obtains through step (2) is specially:

All subwindows of 30 * 30 of intercepting entire image; Each subwindow is successively through 20 grades of gesture feature sorters; Eliminate non-target gesture subwindow step by step, have only through the subwindow of all 20 grades of gesture feature sorters and just confirm as the target gesture, if in this layer search, do not find the target gesture; Then with subwindow with 1.3 times of amplifications, detect through the gesture feature sorter again.

In search procedure, in order to improve search efficiency, if the previous frame image has searched the target gesture, then the hunting zone with current frame image is reduced into the zone around the target gesture region in the previous frame image.As shown in Figure 4, suppose that the center, target gesture region (being the inside casing among Fig. 4) that in the previous frame image, searches is some O, some O is d to the left margin distance of target gesture region ₁, some O is d to the right margin distance ₂, distance is d to some O to the coboundary ₃, some O is at d to the lower boundary distance ₄, rectangular search frame (being the housing among Fig. 4) is made for being the center with an O in then new region of search, and mid point O is 2*d to rectangular search frame left margin distance ₁, O is 2*d to rectangular search frame right margin distance ₂, O is 2*d to rectangular search upper frame edge circle distance ₃, O is 2*d to rectangular search frame lower boundary distance ₄

(5) finger tip location:

(5-1) adopt the moving shape model of the gesture in the step (1) that the target gesture that searches is estimated match, obtain the state space initial value χ of target gesture ₀, make the initial value p of posterior probability density ₀(χ ₀| Z ₀)=1;

(5-2) the target gesture is observed, utilized the particle filter tracking algorithm that the state of target gesture is carried out the iteration renewal through adaptive skin color segmentation method;

Said adaptive skin color segmentation method may further comprise the steps:

(5-2-1) set up TSL space complexion model:

Be the TSL color space with the RGB color space conversion at first through following formula:

$\left\{\begin{array}{c}T=\frac{1}{2\mathrm{\π}}{\tan }^{-1}\left(\frac{r^{\′}}{g^{\′}}\right)+0.5\\ S=\sqrt{\frac{9}{5}(r^{\′2}+g^{\′2})}\\ L=0.299*R+0.587*G+0.114*B\\ \end{array}\right.$

Wherein $r^{\′}=(r-\frac{1}{3}),$ $g^{\′}=(g-\frac{1}{3})$

$r=\frac{R}{R+G+B},$ $g=\frac{G}{R+G+B}$

R, G, B are respectively the RGB component under the rgb color model; T, S, L are respectively the TSL component under the TSL colour model;

(5-2-2) carry out colour of skin filtering: face and hand region through comprising the image of area of skin color to 500 are sampled, the equal value matrix E and the covariance matrix ∑ of the two-dimentional Gaussian distribution probability distribution parameters of T and S under the estimation TSL model;

Each pixel is detected, if the C=that the T of a pixel and S component are formed (T, S) vector is lower than threshold value Threshold with the mahalanobis distance of mean vector E, thinks that then this pixel belongs to area of skin color; Said mahalanobis distance d=(C-E) ^T∑ ^-1(C-E);

Said threshold value Threshold is confirmed by following process:

According to equal value matrix E and covariance matrix ∑ estimation C=(T, S) distance of vector and mean vector E obtains initial value;

Calculate the degree of confidence confidence of each threshold value Threshold according to following formula:

$\mathrm{Confidence}=\frac{\mathrm{PosSkin}}{\mathrm{MaskArea}}*2-\frac{\mathrm{NegSkin}}{\mathrm{BgArea}}$

Template in the complexion model is provided with as shown in Figure 5.PosSkin is meant the quantity of the pixel of the colour of skin in the colour of skin template zone (being gesture zone among Fig. 5) in the following formula; MaskArea is meant the total area in colour of skin template zone; NegSkin is meant the quantity of skin pixel point in the background template zone (being the black background among Fig. 5), and BgArea is meant the area in background template zone.

Getting the maximum threshold value Threshold of degree of confidence is the TSL complexion model parameter of current the best, and upgrades TSL space complexion model;

(5-2-3) bianry image that obtains after the colour of skin filtering is carried out dilation operation, to reduce the cavity that colour of skin filtering causes.

Said iteration renewal process is specific as follows:

The observed quantity that defines the t-1 two field picture is Z _T-1, then the state space of target gesture is χ during the t-1 frame _T-1Posterior probability be p _T-1(χ _T-1| Z _T-1);

The optimal profile state of target gesture during according to the t-1 frame

The estimation probability of the optimal profile state of target gesture is during to the t frame: p _t(χ _t| χ _T-1); Order

The likelihood probability of the profile state of target gesture is p during the t frame _t(Z _t| χ _t);

The state space of target gesture is χ when obtaining the t frame _tPosterior probability be:

p _t(χ _t|Z _t)＝p _t(Z _t|χ _t)*∫p _t(χ _t|χ _t-1)*p _t-1(χ _t-1|Z _t-1)dχ _t-1/p _t(Z _t)；

The posterior probability density p of target gesture during according to the t frame _t(χ _t| Z _t) the optimal profile state of target gesture does when estimating the t frame

(5-3) when the t frame, indicate the coordinate A of forefinger finger tip and the coordinate B of thumb finger tip on the optimal profile of target gesture;

(6) mouse mappings:

As shown in Figure 6, make up right-angle triangle ABC: cross the A point and do vertical line, cross the B point and do horizontal line, intersection point is C, and establishing line segment AB and CB angle is α, and the angle of line segment BA and CA is β, and α is less than threshold alpha _TBe mapped as the left mouse button click action, Fig. 7 clicks synoptic diagram for left mouse button.β is less than threshold value beta _TBe mapped as and click action by mouse right button, Fig. 8 is for clicking synoptic diagram by mouse right button.

The mid point M of line taking section AB is mapped as the relative coordinate of mouse, establishes that the forefinger fingertip location is A in the previous frame image ^*, the coordinate B of thumb finger tip ^*, line segment A ^*B ^*Mid point be M ^*, then directed line segment M*M is mapped as the relatively move vector of mouse on screen, the synoptic diagram that Fig. 9 moves for mouse.

The foregoing description is a preferred implementation of the present invention; But embodiment of the present invention is not limited by the examples; Other any do not deviate from change, the modification done under spirit of the present invention and the principle, substitutes, combination, simplify; All should be the substitute mode of equivalence, be included within protection scope of the present invention.

Claims (8)

1. the gesture mouse recognition methods based on machine vision is characterized in that, may further comprise the steps:

(1) sets up the moving shape model of gesture: at first the profile of the open gesture of the five fingers is carried out the basic gesture modeling of B batten; On the basis of basic gesture model, set up the moving shape model of the gesture of B batten then, shown in the following formula of its state space:

χ＝(x，y，θ，s，λ，θ ₁，l ₁，θ ₂，l ₂，θ ₃，l ₃，θ ₄，l ₄，θ ₅，θ ₆)

Wherein, x, y are respectively the horizontal ordinate and the ordinate of the position of gesture; θ representes the gesture anglec of rotation planar; S represents the size of gesture; θ ₁, l ₁Represent angle and the little finger length of little finger respectively around its root rotation; θ ₂, l ₂The expression third finger is around the angle and the nameless length of its root rotation respectively; θ ₃, l ₃Represent angle and the middle finger length of middle finger respectively around its root rotation; θ ₄, l ₄Represent angle and the forefinger length of forefinger respectively around its root rotation; θ ₆The expression thumb is around the angle of its root rotation; θ ₅The expression thumb is around the angle of joint, center rotation;

(2) gesture is carried out off-line training, obtain the gesture feature sorter;

(3) images acquired;

(4) extract the partial binary mode characteristic of image, the gesture feature sorter ferret out gesture in image that obtains through step (2); When searching the target gesture, carry out step (5);

(5) finger tip location:

(5-1) adopt the moving shape model of the gesture in the step (1) that the target gesture that searches is estimated match, obtain the state space initial value χ of target gesture ₀, make the initial value p of posterior probability density ₀(χ ₀| Z ₀)=1;

(5-2) the target gesture is observed, utilized the particle filter tracking algorithm that the state of target gesture is carried out the iteration renewal, said iteration renewal process is specific as follows:

The observed quantity that defines the t-1 two field picture is Z _T-1, then the state space of target gesture is χ during the t-1 frame _T-1Posterior probability be p _T-1(χ _T-1| Z _T-1);

The optimal profile state of target gesture during according to the t-1 frame

The optimal profile state of target gesture during to the t frame

The estimation probability be: p _t(χ _t| χ _T-1); Order

The likelihood probability of the profile state of target gesture is p during the t frame _t(Z _t| χ _t);

The state space of target gesture is χ during the t frame _tPosterior probability be:

p _t(χ _t|Z _t)＝p _t(Z _t|χ _t)*∫p _t(χ _t|χ _t-1)*p _t-1(χ _t-1|Z _t-1)dχ _t-1/p _t(Z _t)；

The posterior probability density p of target gesture during according to the t frame _t(χ _t| Z _t) the optimal profile state of target gesture does when estimating the t frame

(5-3) when the t frame, indicate the coordinate A of forefinger finger tip and the coordinate B of thumb finger tip on the optimal profile of target gesture;

(6) mouse mappings:

Make up right-angle triangle ABC: cross the A point and do vertical line, cross the B point and do horizontal line, intersection point is C;

If line segment AB and CB angle are α, the angle of line segment BA and CA is β, and α is less than threshold alpha _TBe mapped as the left mouse button click action, β is less than threshold value beta _TBe mapped as and click action by mouse right button;

The mid point M of line taking section AB is mapped as the relative coordinate of mouse; If the forefinger fingertip location is A in the previous frame image ^*, the coordinate B of thumb finger tip ^*, line segment A ^*B ^*Mid point be M ^*, then directed line segment M*M is mapped as the relatively move vector of mouse on screen.

2. the gesture mouse recognition methods based on machine vision according to claim 1 is characterized in that, step (2) is said carries out off-line training to gesture, obtains the gesture feature sorter, is specially:

Use 600 open one's eyes wide the mark gesture as positive sample and 1200 non-target gestures as negative sample; The opencv_traincascade program that use is increased income to be provided among the Flame Image Process class libraries OpenCV is trained.

3. the gesture mouse recognition methods based on machine vision according to claim 1 is characterized in that, the said gesture feature sorter ferret out gesture in image that obtains through step (2) of step (4) is specially:

All subwindows of 30 * 30 of intercepting entire image, each subwindow are eliminated non-target gesture subwindow step by step successively through 20 grades of gesture feature sorters, confirm as the target gesture through the subwindow of all 20 grades of gesture feature sorters; If in this layer search, do not find the target gesture, then with subwindow with 1.3 times of amplifications, detect through the gesture feature sorter again.

4. the gesture mouse recognition methods based on machine vision according to claim 1 and 2; It is characterized in that; In the said ferret out gesture of step (2) process; If the previous frame image has searched the target gesture, then the hunting zone with current frame image is reduced into the zone around the target gesture region in the previous frame image, is specially:

The left margin distance is d to target gesture region if the center of the target gesture region that searches in the previous frame image is for some O, some O ₁, some O is d to the right margin distance ₂, distance is d to some O to the coboundary ₃, some O is d to the lower boundary distance ₄, the rectangular search frame is made for being the center with an O in then new region of search, and some O is 2*d to rectangular search frame left margin distance ₁, some O is 2*d to rectangular search frame right margin distance ₂, some O is 2*d to rectangular search upper frame edge circle distance ₃, some O is 2*d to rectangular search frame lower boundary distance ₄

5. the gesture mouse recognition methods based on machine vision according to claim 1 is characterized in that step (5-2) is said to be observed the target gesture, specifically observes through adaptive skin color segmentation method; Said adaptive skin color segmentation method may further comprise the steps:

(5-2-1) set up TSL space complexion model;

(5-2-2) carry out colour of skin filtering;

(5-2-3) bianry image that obtains after the colour of skin filtering is carried out dilation operation.

6. the gesture mouse recognition methods based on machine vision according to claim 5 is characterized in that, the said TSL of foundation of step (5-2-1) space complexion model is specially:

Is the TSL color space through following formula with the RGB color space conversion:

$\left\{\begin{array}{c}T=\frac{1}{2\mathrm{\π}}{\tan }^{-1}\left(\frac{r^{\′}}{g^{\′}}\right)+0.5\\ S=\sqrt{\frac{9}{5}(r^{\′2}+g^{\′2})}\\ L=0.299*R+0.587*G+0.114*B\\ \end{array}\right.$

Wherein $r^{\′}=(r-\frac{1}{3}),$ $g^{\′}=(g-\frac{1}{3})$

$r=\frac{R}{R+G+B},$ $g=\frac{G}{R+G+B}$

R, G, B are respectively the RGB component under the rgb color model; T, S, L are respectively the TSL component under the TSL colour model.

7. the gesture mouse recognition methods based on machine vision according to claim 6 is characterized in that, step (5-2-2) is said carries out colour of skin filtering, is specially:

500 face and hand region that comprise the image of area of skin color are sampled the equal value matrix E and the covariance matrix ∑ of the two-dimentional Gaussian distribution probability distribution parameters of T and S under the estimation TSL model;

Each pixel is detected, if the C=that the T of a pixel and S component are formed (T, S) vector is lower than threshold value Threshold with the mahalanobis distance of mean vector E, thinks that then this pixel belongs to area of skin color; Said mahalanobis distance d=(C-E) ^T∑ ^-1(C-E).

8. the gesture mouse recognition methods based on machine vision according to claim 7 is characterized in that, said threshold value Threshold is confirmed by following process:

According to equal value matrix E and covariance matrix ∑ estimation C=(T, S) distance of vector and mean vector E obtains initial value;

Calculate the degree of confidence confidence of each threshold value Threshold according to following formula:

$\mathrm{Confidence}=\frac{\mathrm{PosSkin}}{\mathrm{MaskArea}}*2-\frac{\mathrm{NegSkin}}{\mathrm{BgArea}}$

Wherein PosSkin is meant the quantity of the pixel of the colour of skin in the colour of skin template zone, and MaskArea is meant the total area in colour of skin template zone, and NegSkin is meant the quantity of skin pixel point in the background template zone, and BgArea is meant the area in background template zone;

The threshold value Threshold that degree of confidence is maximum is updated to the TSL complexion model parameter of current the best, and upgrades TSL space complexion model.

Images