CN104200199B - Bad steering behavioral value method based on TOF camera - Google Patents

Abstract

The invention discloses a kind of bad steering behavioral value method based on TOF camera, is characterized in that carrying out as follows：Step 1, acquisition depth image；Step 2, the detection of head zone and matching；The detect and track of step 3, hand region.The present invention can accurately and effectively detect the bad steering behavior of driver in the case where driver's normal driving is not disturbed, and improve recognition accuracy, so as to reduce the incidence rate of accident.

Description

Bad steering behavioral value method based on TOF camera

Technical field

The present invention relates to a kind of Motion Recognition and safety assistant driving technology based on TOF camera, belongs to 3D vision Intelligent traffic safety field.

Background technology

With the continuous improvement of Chinese Urbanization and motorization level, traffic safety problem is increasingly serious.Analyze me Road traffic accident reason finds that about 90% vehicle accident is caused due to the bad steering behavior of driver in recent years for state 's.Therefore, just become guarantee road traffic to specifically how the research of driving behavior is intervened to unsafe driving behavior The key problem of safety.

The video image that 2D cameras obtain driver is mainly used to the research of driving behavior both at home and abroad at present, so as to right The action of eyelid, the closure situation of eyes, nod and the external state of the driver such as mouth action is detected, and develop warning Monitoring device is waking up driver.But as 2D cameras are easily affected by illumination, shade, human body complexion etc., while 3D is true Real field scape has the loss of range information when projecting to 2D planes, so meeting band when driving behavior is recognized is being carried out with 2D cameras Carry out the loss of information, the problems such as detection is inaccurate, discrimination is not high.

The content of the invention

The present invention is the weak point for overcoming prior art to exist, it is proposed that a kind of bad steering row based on TOF camera For detection method, effective detection can go out the bad steering behavior of driver in the case where driver's normal driving is not disturbed, carry High recognition accuracy, so as to reduce the incidence rate of accident.

The present invention solves technical problem and adopts the following technical scheme that：

A kind of the characteristics of bad steering behavioral value method based on TOF camera of the present invention is to carry out as follows：

Step 1, acquisition depth image：

Obtained in time period T=(t using TOF camera₁,t₂,...t_v...,t_m) interior driver driving behavior depth Image, chooses t₁The driving behavior depth image at moment is ID image；t₂Moment is to t_mThe driving behavior depth map at moment As being sequence depth image；

Step 2, the detection of head zone and matching：

2.1st, the human face region in the ID image is detected using AdaBoost algorithms, and utilizes head Rectangle frame is labeled to detected human face region, and obtains the center of the head rectangle frame；

2.2nd, the head rectangle frame is outwards expanded into A%, obtains head and expand rectangle frame；

2.3rd, it is made up of the pixel of human face region in the head rectangle frame and refers to three-dimensional point cloud；Expanded by the head In rectangle frame, the pixel of human face region constitutes three-dimensional point cloud to be matched, using ICP algorithm to the three-dimensional point cloud to be matched with The reference three-dimensional point cloud carries out three-dimensional registration, obtains the final iterationses of ICP algorithm, according to the final iterationses with The threshold value of setting is compared, if the threshold value of final iterationses >=set, is judged as bad steering behavior, otherwise Head is judged for normal driving behavior, and continue executing with step 3；

The detect and track of step 3, hand region：

3.1st, the whole of driver in the ID image is labeled above the waist, obtains upper part of the body rectangle frame；

3.2nd, using depth value in the upper part of the body rectangle frame minimum region as hand region, and with boundary rectangle frame pair The hand region is labeled, the width of a width of described head rectangle frame of the boundary rectangle frame, the boundary rectangle frame The twice of the length of a length of head rectangle frame；

3.3rd, the hand region in external rectangle frame in the sequence depth image is carried out using Kalman filtering algorithm Tracking, obtains the center of external rectangle frame in the sequence depth image；

3.4th, obtain European between the center of boundary rectangle frame and the center of the head rectangle frame Distance；

If the 3.5, the distance threshold of the Euclidean distance >=set, is judged as bad steering behavior, is otherwise judged as Normal driving behavior.

The characteristics of bad steering behavioral value method of the present invention based on TOF camera, lies also in：

In the step 2.1, AdaBoost algorithms are the steps of carrying out：

Step a, the definition sequence depth image are training sample I={ (x₁,y₁),(x₂,y₂),…,(x_i,y_i),…, (x_N,y_N), 1≤i≤N, x_iRepresent i-th training sample, y_i∈ { 0,1 }, y_i=0 represents positive sample；y_i=1 negative sample；

Step b, the extraction that Harr features are carried out to each training sample in the training sample I, obtain Harr features Set F={ f₁,f₂,...,f_j,...,f_M, f_jRepresent j-th Harr feature, 1≤j≤M；

Step c, j-th Harr feature f that i-th training sample is obtained using formula (1)_jGrader h_j(x_i)：

In formula (1), p_jFor the grader h_j(x_i) directioin parameter, p_j=± 1, θ is threshold value；

Step d, repeat step c, so as to obtain M grader set F={ h₁(x_i),h₂(x_i),...,h_j(x_i),..., h_M(x_i)}；

Step e, j-th grader h that i-th training sample is obtained using formula (2)_j(x_i) weighting classification error ε_j：

In formula (2), ω_iRepresent the weight after i-th training sample normalization；

Step f, repeat step e, so as to obtain M weighting classification error set E={ ε₁,ε₂,...,ε_j,...,ε_M}；

Step g, the grader for choosing weighting classification error minimum in the weighting classification error set E, with the weighting Harr features corresponding to the minimum grader of error in classification are detected to the human face region in the ID image.

ICP algorithm in the step 2.3 is to carry out as follows：

Step a, definition three dimensions are R³；

The initial coordinate for defining pixel in the three-dimensional point cloud to be matched is { P_i|P_i∈R³, i=1 ...., N_P, N_PTable Show pixel sum in the three-dimensional point cloud to be matched；

The initial coordinate for defining pixel in the reference three-dimensional point cloud is { Q_i|Q_i∈R³, i=1 ...., N_Q}；N_QRepresent State the sum with reference to pixel in three-dimensional point cloud；

Step b, definition iteration total degree are W；

When defining kth time iteration, k ∈ W；Initial coordinate { Q to pixel in the reference three-dimensional point cloud_i|Q_i∈R³,i =1 ...., N_QIn each coordinate points, the coordinate { P of pixel in three-dimensional point cloud to be matched_i ^k|P_i ^k∈R³, i= 1,....,N_PIn all find out distance { Q_i|Q_i∈R³, i=1 ...., N_QIn the nearest point of each coordinate points, by the distance Nearest point composition refers to the coordinate of pixel in three-dimensional point cloud

When step c, definition+1 iteration of kth, pixel in three-dimensional point cloud to be matched is obtained respectively using formula (3) and formula (4) The coordinate of point is { P_i ^k+1|P_i ^k+1∈R³, i=1 ...., N_PAnd with reference to the coordinate of pixel in three-dimensional point cloud be

In formula (3) and formula (4), R_O ^k+1Represent three-dimensional rotation matrix；t^k+1Represent translation vector；

Step d, using formula (5) obtain+1 iteration of kth three-dimensional point cloud to be matched in pixel coordinate { P_i ^k+1And The coordinate of pixel in the reference three-dimensional point cloud of kth time iterationBetween Euclidean distance d^k+1：

Step e, repeat step b- step d, obtain the Euclidean distance d^k+1Minima when, stop iteration and acquisition most Whole iterationses W'.

Kalman filtering algorithm in the step 3.3 is to carry out as follows：

1), make t₂The hand state at moment isInitialization

2) t in the boundary rectangle frame of the sequence depth image, is obtained using formula (6)_vThe hand predicted state at moment

In formula (6),Hand state transfer matrix is represented,T is between the two neighboring moment Interval；Represent in the sequence depth image t in external rectangle frame_v-1The hand state at moment；

3), t is obtained using formula (7)_v-1The hand state of the boundary rectangle inframe at momentCovariance matrix

In formula (7),Represent t in the boundary rectangle frame of the sequence depth image_v-1The hand predicted state at moment；Represent t in the boundary rectangle frame of the sequence depth image_v-2The hand state at moment；

4), the t is obtained using formula (8)_vThe hand predicted state at momentCovariance matrix

In formula (8),For hand state transfer matrixTransposition；Dynamic noise covariance matrix is represented, and is obeyed Standard normal distribution

5), t is updated using formula (9)_vThe hand state of the boundary rectangle inframe at moment

In formula (9),For hand state observing matrix,For t_vMoment Kalman filtering gain Coefficient；

6), Kalman filtering gain coefficient is calculated using formula (10)

In formula (10),For hand state observing matrixTransposition,Noise covariance matrix is represented, and obeys mark Quasi normal distribution R_k～N (0,1)；

7), repeat step 1) -6), constantly update t_vThe hand state of the boundary rectangle inframe at momentSo as to realize The tracking of boundary rectangle inframe hand region, by external rectangle frame in the hand region acquisition sequence depth image Heart position.

Three-dimensional rotation matrix R in step c_O ^k+1With translation vector t^k+1Obtain as follows：

1) coordinate { P of pixel in the three-dimensional point cloud to be matched when+1 iteration of kth, is obtained using formula (11) and (12)_i ^{k +1}Barycenter μ_PWith the coordinate with reference to pixel in three-dimensional point cloudBarycenter μ_Q：

2), obtain the coordinate { P of pixel in the three-dimensional point cloud to be matched using formula (13) and (14) respectively_i ^k+1Relative In barycenter μ_PTranslationWith the coordinate of pixel in the reference three-dimensional point cloudRelative to barycenter μ_QTranslation

3), the translation is obtained using formula (15)And translationBetween correlation matrix K_αβ：

In formula (15), α, β=1,2,3；

4), correlation matrix K is obtained using formula (16)_αβThe four-dimensional symmetrical matrix K for being constructed：

5), Maximum characteristic root is obtained using the four-dimensional symmetrical matrix K；And it is special to obtain unit according to the Maximum characteristic root Levy vectorial q=[q₀,q₁,q₂,q₃]^T；

6), antisymmetric matrix K (q) is obtained using formula (17)：

7), three-dimensional rotation matrix is obtained using formula (18)

8), translation vector t is obtained using formula (19)^k+1：

Compared with prior art, beneficial effects of the present invention are embodied in：

1st, the present invention realizes driver head by AdaBoost algorithms and Kalman filtering and detects and hand state Tracking, and the three-dimensional point cloud registration of current depth figure and reference depth figure is completed with ICP algorithm, solve of the prior art The problem that 2D cameras are easily affected by worn dress ornament color of illumination, shade, human body complexion and driver etc., it is bad so as to improve The accuracy of detection of driving behavior, reduces accident rate；

2nd, driver recognition methodss of the present invention using TOF range informations, because TOF camera obtains depth image speed soon, Frame frequency up to 40fps, with real-time it is good the characteristics of；

3rd, AdaBoost algorithms of the present invention are a kind of high-precision graders, will not be transported because of driver head The dynamic detection error brought and tremendous influence is produced to subsequent detection result, and because the framework of algorithm is simple, it is special that it goes without doing Screening is levied, so as to solve the speed issue of detection, with preferable recognition effect.

4th, ICP algorithm of the present invention need not be split to pending point set and feature extraction, and select Accurately in the case of initial position, good Algorithm Convergence can be obtained, can be obtained in terms of the registration of three-dimensional point cloud Point-device registration effect.

5th, Kalman filtering algorithm of the present invention is the optimum criterion with least mean-square error to estimate, in mathematics It is fairly simple on framework, and be optimum linearity recursive filtering method, energy effectively solving is based on irregular hand continuity chart picture Target following, have the advantages that amount of calculation is little.

Description of the drawings

Fig. 1 detecting system schematic diagrams of the present invention；

Fig. 2 detection method flow charts；

Label in figure：1 head rectangle frame；2 heads expand rectangle frame；3 upper part of the body rectangle frames；4 boundary rectangle frames；5TOF phases Machine.

Specific embodiment

In the present embodiment, a kind of bad steering behavioral value method based on TOF camera is to be driven by installation in the car The TOF camera of member oblique upper obtains the depth image of driver in real time, and by conjecture face rectangle frame, using ICP, (iteration is nearest Point) present frame depth image and reference picture carry out three-dimensional registering by algorithm, judge whether to there occurs according to iterationses bad Driving behavior；Hand is detected using background subtraction and is tracked, sentenced with conjecture face rectangle frame position according to hand position It is disconnected whether to there occurs bad steering behavior, particularly carry out according to the following steps：

Step 1, acquisition depth image：

TOF camera is arranged on the oblique upper for driving indoor driver, and connects a microprocessor peace that can be reported to the police It is mounted on vehicle console, is obtained in time period T=(t using TOF camera₁,t₂,...t_v...,t_m) interior driver driving row For depth image, choose t₁The driving behavior depth image at moment is ID image；t₂Moment is to t_mThe driving row at moment It is sequence depth image for depth image；

Step 2, the detection of head zone and matching：

2.1st, the human face region in ID image is detected using AdaBoost algorithms, and utilizes head rectangle Frame is labeled to detected human face region, and obtains the center of head rectangle frame；

Specifically, AdaBoost algorithms are carried out as follows：

A), defined nucleotide sequence depth image is training sample I={ (x₁,y₁),(x₂,y₂),…,(x_i,y_i),…,(x_N,y_N), 1≤i≤N, x_iRepresent i-th training sample, y_i∈ { 0,1 }, y_i=0 represents positive sample；y_i=1 negative sample；

B) extraction of Harr features, is carried out to each training sample in training sample I, Harr characteristic set F=are obtained {f₁,f₂,...,f_j,...,f_M, f_jRepresent j-th Harr feature, 1≤j≤M；

C) j-th Harr feature f of i-th training sample, is obtained using formula (1)_jGrader h_j(x_i)：

In formula (1), p_jFor grader h_j(x_i) directioin parameter, p_j=± 1, θ is threshold value, is taken as 0.15 in this example；

D), repeat step c), so as to obtain M grader set F={ h₁(x_i),h₂(x_i),...,h_j(x_i),...,h_M (x_i)}；

E) j-th grader h of i-th training sample, is obtained using formula (2)_j(x_i) weighting classification error ε_j：

In formula (2), ω_iRepresent the weight after i-th training sample normalization；

F), repeat step e), so as to obtain M weighting classification error set E={ ε₁,ε₂,...,ε_j,...,ε_M}；

G) the minimum grader of weighting classification error in weighting classification error set E, is chosen, it is minimum with weighting classification error Grader corresponding to Harr features the human face region in ID image is detected.

2.2nd, head rectangle frame is outwards expanded into A%, here expand the reason for be driver in normal driving, have micro- Little head movement, not belongs to bad steering behavior within this range, and A takes 10 in this example, obtains head and expands rectangle Frame；

2.3rd, it is made up of the pixel of human face region in head rectangle frame and refers to three-dimensional point cloud；Wherein, three-dimensional point cloud is one The set of group pixel, the pixel for expanding human face region in rectangle frame by head constitute three-dimensional point cloud to be matched, are calculated using ICP Method, ICP algorithm are iterative closest point algorithm, to three-dimensional point cloud to be matched with three-dimensional registering, acquisition ICP is carried out with reference to three-dimensional point cloud The final iterationses of algorithm, are compared according to threshold value of the final iterationses with setting, if final iterationses >=set Fixed threshold value, set threshold value is 50 in this example, then be judged as bad steering behavior, otherwise judge head normally to drive Sail behavior；, and continue executing with step 3；The foundation judged here according to iterationses is that ICP algorithm is a kind of iteration With algorithm, need in the matching process to carry out ceaselessly iteration, the final iteration if the matching degree height of two width driving behavior images Number of times is few, is now normal driving, and if the big matching degree of two width driving behavior image difference is low, final iterationses are more, now For bad steering, so selecting according to the matching iterationses of driver's driving behavior image to judge whether driver there occurs Bad steering.

Specifically, ICP algorithm is to carry out as follows：

H) it is R, to define three dimensions³；

The initial coordinate for defining pixel in three-dimensional point cloud to be matched is { P_i|P_i∈R³, i=1 ...., N_P, N_PExpression is treated Pixel sum in matching three-dimensional point cloud；

Definition is { Q with reference to the initial coordinate of pixel in three-dimensional point cloud_i|Q_i∈R³, i=1 ...., N_Q}；N_QGinseng is stated in expression Examine the sum of pixel in three-dimensional point cloud；

I) it is W, to define iteration total degree；

When defining kth time iteration, k ∈ W；To the initial coordinate { Q with reference to pixel in three-dimensional point cloud_i|Q_i∈R³, i= 1,....,N_QIn each coordinate points, the coordinate { P of pixel in three-dimensional point cloud to be matched_i ^k|P_i ^k∈R³, i= 1,....,N_PIn all find out away from { Q_i|Q_i∈R³, i=1 ...., N_QIn the nearest point of each coordinate points, these nearest points Composition refers to the coordinate of pixel in three-dimensional point cloud

When j), defining+1 iteration of kth, pixel in three-dimensional point cloud to be matched is obtained respectively using formula (3) and formula (4) Coordinate is { P_i ^k+1|P_i ^k+1∈R³, i=1 ...., N_PAnd with reference to the coordinate of pixel in three-dimensional point cloud be

In formula (3) and formula (4), R_O ^k+1Represent three-dimensional rotation matrix；t^k+1Represent translation vector；

Specifically, three-dimensional rotation matrix R_O ^k+1With translation vector t^k+1Obtain as follows：

Coordinate { the P of pixel in three-dimensional point cloud to be matched when+1 iteration of kth j1), is obtained using formula (5) and (6)_i ^{k +1}Barycenter μ_PWith the coordinate with reference to pixel in three-dimensional point cloudBarycenter μ_Q：

J2), obtain the coordinate { P of pixel in three-dimensional point cloud to be matched using formula (7) and (8) respectively_i ^k+1Relative to matter Heart μ_PTranslationWith the coordinate with reference to pixel in three-dimensional point cloudRelative to barycenter μ_QTranslation

J3), translated using formula (9)And translationBetween correlation matrix K_αβ：

In formula (9), α, β=1,2,3；

J4), correlation matrix K is obtained using formula (10)_αβThe four-dimensional symmetrical matrix K for being constructed：

J5), Maximum characteristic root is obtained using four-dimensional symmetrical matrix K；And unit character vector q is obtained according to Maximum characteristic root =[q₀,q₁,q₂,q₃]^T；

J6), antisymmetric matrix K (q) is obtained using formula (11)：

J7), three-dimensional rotation matrix is obtained using formula (12)

J8), translation vector t is obtained using formula (13)^k+1：

K) coordinate { P of pixel in the three-dimensional point cloud to be matched of+1 iteration of kth, is obtained using formula (14)_i ^k+1And kth The coordinate of pixel in the reference three-dimensional point cloud of secondary iterationBetween Euclidean distance d^k+1：

L), repeat step i)-step k), obtains Euclidean distance d^k+1Minima when, stop iteration and simultaneously obtain final iteration Number of times W'.

The detect and track of step 3, hand region：

3.1st, the whole of driver in ID image is labeled above the waist, obtains upper part of the body rectangle frame；

3.2nd, using the minimum region of depth value in upper part of the body rectangle frame as hand region, because hand is apart from TOF camera Closely, so its depth value is minimum, and hand region is labeled with boundary rectangle frame, a width of head rectangle of boundary rectangle frame The width of frame, the twice of the length of a length of head rectangle frame of boundary rectangle frame；Because in general, the hand and head size of people are poor It is few the same, it is contemplated that during normal driving, hand has certain range of activity, so being set to the width of boundary rectangle frame here The width of head rectangle frame, the length of boundary rectangle frame are set to the twice of the length of head rectangle frame.

3.3rd, the hand region in external rectangle frame in sequence depth image is tracked using Kalman filtering algorithm, Obtain the center of external rectangle frame in sequence depth image；The center of boundary rectangle frame is obtained according to geometrical relationship.

Specifically, Kalman filtering algorithm is to carry out as follows：

M), make t₂The hand state at moment isInitialization

N) t in the boundary rectangle frame of sequence depth image, is obtained using formula (15)_vThe hand predicted state at moment

In formula (15),Hand state transfer matrix is represented,T is between the two neighboring moment Interval；T takes 0.2s in this example,T in external rectangle frame in expression sequence depth image_v-1The hand shape at moment State；

O), t is obtained using formula (16)_v-1The hand state of the boundary rectangle inframe at momentCovariance matrix

In formula (16),Represent t in the boundary rectangle frame of sequence depth image_v-1The hand predicted state at moment；Represent t in the boundary rectangle frame of sequence depth image_v-2The hand state at moment；

P), t is obtained using formula (17)_vThe hand predicted state at momentCovariance matrix

In formula (17),For hand state transfer matrixTransposition；Dynamic noise covariance matrix is represented, and is taken From standard normal distribution

Q), t is updated using formula (18)_vThe hand state of the boundary rectangle inframe at moment

In formula (18),For hand state observing matrix,For t_vMoment Kalman filtering increases Beneficial coefficient；

R), Kalman filtering gain coefficient is calculated using formula (19)

In formula (19),For hand state observing matrixTransposition,Noise covariance matrix is represented, and obeys mark Quasi normal distribution R_k～N (0,1)；

S), repeat step m)-q), constantly update t_vThe hand state of the boundary rectangle inframe at momentSo as to realize The tracking of boundary rectangle inframe hand region, obtains the center of external rectangle frame in sequence depth image by hand region.

3.4th, obtain the Euclidean distance between the center and the center of head rectangle frame of boundary rectangle frame；

If the 3.5, distance threshold of Euclidean distance >=set, set distance threshold is 40cm in this example, then It is judged as bad steering behavior, is otherwise judged as normal driving behavior.

Claims (5)

1. a kind of bad steering behavioral value method based on TOF camera, is characterized in that carrying out as follows：

Step 1, acquisition depth image：

Obtained in time period T=(t using TOF camera₁,t₂,...t_v...,t_m) interior driver driving behavior depth image, Choose t₁The driving behavior depth image at moment is ID image；t₂Moment is to t_mThe driving behavior depth image at moment is Sequence depth image；

Step 2, the detection of head zone and matching：

2.1st, the human face region in the ID image is detected using AdaBoost algorithms, and utilizes head rectangle Frame is labeled to detected human face region, and obtains the center of the head rectangle frame；

2.2nd, the head rectangle frame is outwards expanded into A%, obtains head and expand rectangle frame；

2.3rd, it is made up of the pixel of human face region in the head rectangle frame and refers to three-dimensional point cloud；Rectangle is expanded by the head In frame, the pixel of human face region constitutes three-dimensional point cloud to be matched, using ICP algorithm to the three-dimensional point cloud to be matched with it is described Three-dimensional registration is carried out with reference to three-dimensional point cloud, the final iterationses of ICP algorithm are obtained, according to the final iterationses and setting Threshold value be compared, if the threshold value of final iterationses >=set, is judged as bad steering behavior, otherwise judges Head is normal driving behavior, and continues executing with step 3；

The detect and track of step 3, hand region：

3.1st, the whole of driver in the ID image is labeled above the waist, obtains upper part of the body rectangle frame；

3.2nd, using the minimum region of depth value in the upper part of the body rectangle frame as hand region, and with boundary rectangle frame to described Hand region is labeled, the width of a width of described head rectangle frame of the boundary rectangle frame, the boundary rectangle frame it is a length of The twice of the length of the head rectangle frame；

3.3rd, the hand region in external rectangle frame in the sequence depth image is tracked using Kalman filtering algorithm, Obtain the center of external rectangle frame in the sequence depth image；

3.4th, obtain the Euclidean distance between the center of boundary rectangle frame and the center of the head rectangle frame；

If the 3.5, the distance threshold of the Euclidean distance >=set, is judged as bad steering behavior, otherwise it is judged as normal Driving behavior.

2. the bad steering behavioral value method based on TOF camera according to claim 1, is characterized in that：The step In 2.1, AdaBoost algorithms are the steps of carrying out：

Step a, the definition sequence depth image are training sample I={ (x₁,y₁),(x₂,y₂),…,(x_i,y_i),…,(x_N, y_N), 1≤i≤N, x_iRepresent i-th training sample, y_i∈ { 0,1 }, y_i=0 represents positive sample；y_i=1 negative sample；

Step b, the extraction that Harr features are carried out to each training sample in the training sample I, obtain Harr characteristic set F ={ f₁,f₂,...,f_j,...,f_M, f_jRepresent j-th Harr feature, 1≤j≤M；

Step c, j-th Harr feature f that i-th training sample is obtained using formula (1)_jGrader h_j(x_i)：

In formula (1), p_jFor the grader h_j(x_i) directioin parameter, p_j=± 1, θ is threshold value；

Step d, repeat step c, so as to obtain M grader set F={ h₁(x_i),h₂(x_i),...,h_j(x_i),...,h_M (x_i)}；

Step e, j-th grader h that i-th training sample is obtained using formula (2)_j(x_i) weighting classification error ε_j：

${\mathrm{\ϵ}}_j=\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\ω}}_i|h_j\left(x_i\right)-y_i|---\left(2\right)$

In formula (2), ω_iRepresent the weight after i-th training sample normalization；

Step f, repeat step e, so as to obtain M weighting classification error set E={ ε₁,ε₂,...,ε_j,...,ε_M}；

Step g, the grader for choosing weighting classification error minimum in the weighting classification error set E, with the weighting classification Harr features corresponding to the minimum grader of error are detected to the human face region in the ID image.

3. the bad steering behavioral value method based on TOF camera according to claim 1, is characterized in that：The step ICP algorithm in 2.3 is to carry out as follows：

Step a, definition three dimensions are R³；

The initial coordinate for defining pixel in the three-dimensional point cloud to be matched is { P_i|P_i∈R³, i=1 ...., N_P, N_PRepresent institute State pixel sum in three-dimensional point cloud to be matched；

The initial coordinate for defining pixel in the reference three-dimensional point cloud is { Q_i|Q_i∈R³, i=1 ...., N_Q}；N_QGinseng is stated in expression Examine the sum of pixel in three-dimensional point cloud；

Step b, definition iteration total degree are W；

When defining kth time iteration, k ∈ W；Initial coordinate { Q to pixel in the reference three-dimensional point cloud_i|Q_i∈R³, i= 1,....,N_QIn each coordinate points, the coordinate { P of pixel in three-dimensional point cloud to be matched_i ^k|P_i ^k∈R³, i= 1,....,N_PIn all find out distance { Q_i|Q_i∈R³, i=1 ...., N_QIn the nearest point of each coordinate points, by the distance Nearest point composition refers to the coordinate of pixel in three-dimensional point cloud

When step c, definition+1 iteration of kth, pixel in three-dimensional point cloud to be matched is obtained respectively using formula (3) and formula (4) Coordinate is { P_i ^k+1|P_i ^k+1∈R³, i=1 ...., N_PAnd with reference to the coordinate of pixel in three-dimensional point cloud be

$P_i^{k+1}=R_O^k{P}_i^k+t^k---\left(3\right)$

$Q_i^{k+1}=R_O^k{Q}_i^k+t^k---\left(4\right)$

In formula (3) and formula (4), R_O ^k+1Represent three-dimensional rotation matrix；t^k+1Represent translation vector；

Step d, using formula (5) obtain+1 iteration of kth three-dimensional point cloud to be matched in pixel coordinate { P_i ^k+1And kth time The coordinate of pixel in the reference three-dimensional point cloud of iterationBetween Euclidean distance d^k+1：

$d^{k+1}=\frac{1}{N_P}\underset{i=1}{\overset{N_P}{\Σ}}\left|\right|P_i^{k+1}-Q_i^k|{|}^2---\left(5\right)$

Step e, repeat step b- step d, obtain the Euclidean distance d^k+1Minima when, stop iteration and obtain finally changing Generation number W'.

4. the bad steering behavioral value method based on TOF camera according to claim 1, is characterized in that：The step Kalman filtering algorithm in 3.3 is to carry out as follows：

1), make t₂The hand state at moment isInitialization

2) t in the boundary rectangle frame of the sequence depth image, is obtained using formula (6)_vThe hand predicted state at moment

$X_{t_v}^{\′}=A_{t_v}{X}_{t_{v-1}}---\left(6\right)$

In formula (6),Hand state transfer matrix is represented,T be between the two neighboring moment between Every；Represent in the sequence depth image t in external rectangle frame_v-1The hand state at moment；

3), t is obtained using formula (7)_v-1The hand state of the boundary rectangle inframe at momentCovariance matrix

$P_{t_{v-1}}=\mathrm{cov}(X_{t_{v-1}}^{\′}-X_{t_{v-2}})---\left(7\right)$

In formula (7),Represent t in the boundary rectangle frame of the sequence depth image_v-1The hand predicted state at moment； Represent t in the boundary rectangle frame of the sequence depth image_v-2The hand state at moment；

4), the t is obtained using formula (8)_vThe hand predicted state at momentCovariance matrix

$P_{t_v}^{\′}=A_{t_v}{P}_{t_{v-1}}{A}_{t_v}^T+B_{t_{v-1}}---\left(8\right)$

In formula (8),For hand state transfer matrixTransposition；Dynamic noise covariance matrix is represented, and obeys standard Normal distribution

5), t is updated using formula (9)_vThe hand state of the boundary rectangle inframe at moment

$X_{t_v}=X_{t_v}^{\′}+K_{t_v}{H}_{t_v}{X}_{t_v}^{\′}---\left(9\right)$

In formula (9),For hand state observing matrix,For t_vMoment Kalman filtering gain system Number；

6), Kalman filtering gain coefficient is calculated using formula (10)

$K_{t_v}=P_{t_v}^{\′}{H}_{t_v}^T{(H_{t_v}{P}_{t_v}^{\′}{H_{t_v}}^T+R_{t_v})}^{-1}---\left(10\right)$

In formula (10),For hand state observing matrixTransposition,Noise covariance matrix is represented, and just obeys standard State is distributed R_k～N (0,1)；

7), repeat step 1) -6), constantly update t_vThe hand state of the boundary rectangle inframe at momentIt is external so as to realize The tracking of rectangle inframe hand region, obtains the centre bit of external rectangle frame in the sequence depth image by the hand region Put.

5. the bad steering behavioral value method based on TOF camera according to claim 3, is characterized in that：Step c Middle three-dimensional rotation matrix R_O ^k+1With translation vector t^k+1Obtain as follows：

1) coordinate { P of pixel in the three-dimensional point cloud to be matched when+1 iteration of kth, is obtained using formula (11) and (12)_i ^k+1 Barycenter μ_PWith the coordinate with reference to pixel in three-dimensional point cloudBarycenter μ_Q：

${\mathrm{\μ}}_P=\frac{1}{N_P}\underset{i=1}{\overset{N_P}{\Σ}}{P}_i^{k+1}---\left(11\right)$

${\mathrm{\μ}}_Q=\frac{1}{N_T}\underset{i=1}{\overset{N_Q}{\Σ}}{Q}_i^{k+1}---\left(12\right)$

2), obtain the coordinate { P of pixel in the three-dimensional point cloud to be matched using formula (13) and (14) respectively_i ^k+1Relative to matter Heart μ_PTranslationWith the coordinate of pixel in the reference three-dimensional point cloudRelative to barycenter μ_QTranslation

$m_i^P=P_i^{k+1}-{\mathrm{\μ}}_P---\left(13\right)$

$m_i^Q=Q_i^{k+1}-{\mathrm{\μ}}_Q---\left(14\right)$

3), the translation is obtained using formula (15)And translationBetween correlation matrix K_αβ：

$K_{\mathrm{\α}\mathrm{\β}}=\frac{1}{N}\underset{i=1}{\overset{N_P}{\Σ}}{m}_i^P{\left(m_i^Q\right)}^T---\left(15\right)$

In formula (15), α, β=1,2,3；

4), correlation matrix K is obtained using formula (16)_αβThe four-dimensional symmetrical matrix K for being constructed：

$K=\left[\begin{array}{cccc}{K}_{11}+K_{12}+K_{13}& K_{32}-K_{23}& K_{13}-K_{31}& K_{21}-K_{12}\\ K_{32}-K_{23}& K_{11}-K_{22}-K_{33}& K_{12}+K_{21}& K_{13}+K_{31}\\ K_{13}-K_{31}& K_{12}+K_{21}& -K_{11}+K_{22}-K_{33}& K_{23}+K_{32}\\ K_{21}-K_{12}& K_{31}+K_{13}& K_{32}+K_{23}& -K_{11}-K_{22}+K_{33}\end{array}\right]---\left(16\right)$

5), Maximum characteristic root is obtained using the four-dimensional symmetrical matrix K；And according to the Maximum characteristic root obtain unit character to Amount q=[q₀,q₁,q₂,q₃]^T；

6), antisymmetric matrix K (q) is obtained using formula (17)：

$K\left(q\right)=\left[\begin{array}{ccc}0& -q_2& q_1\\ q_2& 0& -q_0\\ -q_1& q_0& 0\end{array}\right]---\left(17\right)$

7), three-dimensional rotation matrix is obtained using formula (18)

$R_O^{k+1}=(q_3^2+q^{T}q+2q_{3}K(q))---\left(18\right)$

8), translation vector t is obtained using formula (19)^k+1：

$t^{k+1}={\mathrm{\μ}}_Q-R_O^{k+1}{\mathrm{\μ}}_P---\left(19\right).$