CN106127776A - Based on multiple features space-time context robot target identification and motion decision method - Google Patents

Abstract

Based on multiple features space-time context robot target identification and motion decision method, belong to robot field.This method carries out Block Cluster first with color and the textural characteristics of image to image, completes the initialization of the space-time context model of the first frame target, and constructs the rarefaction representation equation of object block.Then with cluster block as ultimate unit, the cluster the value of the confidence in cluster process is combined with contextual feature and sets up the confidence map with image block as ultimate unit.Finally, confidence map likelihood probability maximum is the target location of next frame of prediction.Compared with former method, present invention enhances clarification of objective and describe, improve target robustness under complex background, it addition, propose the ideological guarantee of the addition Block Cluster real-time of algorithm.After algorithm completes the recognition and tracking of target, to follow the tracks of the result basis as robot motion's decision-making, complete the robot recognition and tracking to target.

Description

Based on multiple features space-time context robot target identification and motion decision method

Technical field

The invention belongs to robot field, be robot target identification based on multiple features space-time context and motion decision-making Method.

Background technology

Along with the range of application of robot is more and more extensive, the correlation technique of intelligent robot receives attention greatly. Particularly intellect service robot is a popular direction of robot development in recent years, and Intelligent Service machine people is right for service As providing service, first have to carry out the identification of target and carry out the motion decision-making of robot accordingly.Target recognition is machine vision The classical part of area research, domestic and international research worker proposes the tracking of many.Mean Shift algorithm be identify with The classic algorithm of track, for the deficiency of color histogram graph expression color of object distribution in this algorithm, Li proposes a kind of based on poly- The adaptive color rectangular histogram of class algorithm, it is achieved that good recognition and tracking effect.For reducing illumination variation to the process of tracking Impact, Lu et al. utilizes local binary textural characteristics to realize the recognition and tracking to target.But the description that single features is to target Not comprehensive, the poor robustness of tracking, for this problem, Nigam proposes based on particle filter frame fusion gradient and color Feature carries out target recognition tracking, Gu et al. have studied the multiple features fusion target recognition of self adaptation additivity and the property taken advantage of variable weight with Track algorithm.In addition, also have many target recognition track algorithms, but these algorithms become at reply target occlusion and environment All not can solve when changing violent situation.On this problem, Yang et al. proposes to utilize auxiliary mark or target office The solution that portion's contextual information assists target following to be this difficult problem provides new thinking, and Zhang et al. is based on Bayes Framework utilizes spatio-temporal context information more to newly arrive and realizes target following, reply block and during illumination variation situation performance good, But this algorithm only employs the single features of image, quickly move and robustness in the case of background acute variation in target Difference.

Summary of the invention

For with present on problem, it is upper and lower that the present invention proposes a kind of space-time merging multiple features based on space-time context The method for tracking target that literary composition piecemeal sparsity structure represents, i.e. color and textural characteristics first with image carry out piecemeal to image and gather Class, completes the initialization of the space-time context model of the first frame target, and constructs the rarefaction representation equation of object block.Then with poly- Class block is ultimate unit, and the foundation that the cluster the value of the confidence in cluster process combined with contextual feature is with image block for the most single The confidence map of position.Finally, confidence map likelihood probability maximum is the target location of next frame of prediction.With former method phase Ratio, present invention enhances clarification of objective and describes, improve target robustness under complex background, divides it addition, propose to add The ideological guarantee real-time of algorithm of block cluster.After algorithm completes the recognition and tracking of target, to follow the tracks of result as machine The basis of people's motion decision-making, completes the robot recognition and tracking to target.

The technical solution used in the present invention is as follows:

First, carry out scene image segmentation, use simple linear Iterative Clustering (SLIC), utilize image color, Texture and range information carry out cluster and form image block pixel, then give different weights to each image block and screen, Obtain the rarefaction representation equation of image block, the most just complete segmentation and the screening of image.Below to split after picture segmentation To image block be that ultimate unit sets up space-time context model, in the case of known to present frame target location (the first frame by Be manually specified), first obtain the context prior model of image block, be then based under Bayesian frame by target location confidence map with The spatial context model to current frame image of the context prior model of image block, then carries out spatial context and updates To the space-time context model of next frame image, and carry out the renewal of scale parameter.Obtain the space-time context model of next frame After, the confidence of target location is obtained under treatment during a frame by the context prior model of its image block and space-time context model Figure, is i.e. considered the center of target at confidence map maximum probability, the tracking of target the most just completes.After obtaining target location Again its space-time context model is updated, circulates the purpose reaching real-time tracking target with this.Finally obtain mesh in tracking Behind target position, carry out the motion decision-making of robot according to the position of target, it is achieved the robot recognition and tracking to target.Specifically Comprise the following steps:

Step 1, scene image is split

Utilize the color of image, texture and distance similarity feature that pixel is clustered, use the three of Lab color space Dimension colouring information and positional information, introduce local entropy and represent the feature of pixel, and formation compactedness is good, border laminating degree high, shape The image block of shape rule, it is achieved the cluster to pixel.

Step 1.1, obtains local entropy

Local entropy h_iBy formula (1) approximate representation:

Wherein p_iRepresent that current pixel i accounts for the probability of the total number of local pixel.

Step 1.2, clusters pixel

Set i from initial cluster center C_k=[a_k,b_k,h_k,x_k,y_k] start sampling, for reducing sub pixel misspecification Probability and edge noise interference, it is the neighborhood of 3x3 that cluster centre is limited to minimum gradient amount, and each pixel is with recently Adjacent cluster centre distance D_iCluster, such as formula (2):

Wherein a_k,b_k, for the Lab color space [L of pixel k_k,a_k,b_k] color component, h_iFor the office required by formula (1) Portion's entropy, [x_i,y_i] it is the transverse and longitudinal coordinate of pixel i, μ is weight empirical parameter, takes μ=0.4 here.

Step 1.3, updates cluster centre

Once pixel cluster is to closest to center pixel, then update cluster centre Φ_kReplace all pictures in cluster areas The average vector of element:

Wherein, Z_iRepresent with Φ_kCentered by cluster areas, the number of pixels that N is comprised in representing region, C_iRepresent poly- Class center, S_iFor two-dimensional spatial location.

Step 2, image block based on rarefaction representation screens

Introduce rarefaction representation and each piecemeal is given different weight, and as in target after arest neighbors region is screened Context area block, it is possible to preferably tackle the situation of target occlusion and improve real-time.

Step 2.1, sets up the sparse equation of image block, and composes weight

For the n-th piecemeal y_n, the rarefaction representation equation of Weight is:

Wherein, cluster centre distance D_nCalculated by formula (2), S_nFor two-dimensional spatial location.A_nRepresent the sparse system of n-th piece Number vector, W_nIt is sparse weight coefficient,Being the average of different piecemeal sparse coefficient in context area, it is normal that η, μ are positive number Amount, for the impact on rarefaction representation of the specification similarity constraint.

Step 2.2, solves sparse equation

Solve sparse equation and be A_nAnd W_nIt is an optimization problem, solves by iterative method.Main thought assumes that W_nFor fixed value, update A_nValue, then with the A calculated_nValue is fixed value, then solves W_n, constantly repeat, until A_nAnd W_nReceive Hold back to Local Minimum or reach iterations stop value.Assume weight W of all piecemeals₁,W₂,…,W_mIt is it is known that by formula (4) sparse coefficient can be obtained:

Wherein, P_n=(D^TD+Λ(ηW_n))^-1,M_n=P_nD_n ^Ty_n,D =Φ_kΛ is unit diagonal matrix.Piecemeal y is obtained by above formula_nSparse coefficient vector A_nAfter, weight is obtained by following formula:

Can obtain:

Wherein, L is Lagrange multiplier,

Iterations limits and is set to t_max, use sparse coefficient A of solution by iterative method the n-th piecemeal_nWith weight W_nStep Can be summarized as follows:

Input: (1) D_i,(2)y_i

1) initialize, W_n=1, n=1,2 ..., m

2)

3) A is calculated by formula (11)_n

4) W is calculated by formula (13)_n

5)end while

Output: (1) A_n, (2) W_n

Step 3, sets up space-time context model

To complete segmentation and the rarefaction screening of image in above-mentioned steps, obtain with picture block as ultimate unit Sparse graph picture, will be that the first frame is directly specified below in the case of known present frame i.e. t frame i.e. target location, set up based on The context model of image block, carries out target following for next frame i.e. t+1 frame and prepares.

Step 3.1, sets up the contextual feature of target area cluster centre

Assume that the context area of target is divided into M_cIndividual block, uses R_nD () represents target context region in n-th frame The d image block, its cluster centre position is set to CR (d), and visual signature is used insteadRepresenting, the context of cluster centre is special The definition levied is as follows:

Step 3.2, sets up the context prior model of image block

By the visual signature of image block is weighted, set up context prior model based on image block:

Wherein,w_σFor the weighting function of vision noticing mechanism, this parameter is based on figure As the cluster centre of block determines to the distance of target's center position, the image block that the current location of distance objective is the nearest then gives Greater weight, its contextual information is more important to the prediction of next frame target location.

Step 3.3, it is thus achieved that spatial context model based on image block

Now to obtain the present frame i.e. target location confidence map of t frame and context prior model, by Bayesian frame The spatial context model of present frame can be obtained, for improving the calculating time, carry out FFT computing acceleration, then t frame based on image block Spatial context model be:

Wherein F^-1() represents Fourier inversion, and F () represents Fourier transformation.

Step 3.4, space-time updating context

It is weighted adding up to spatial context, obtains carrying out for next frame (t+1 frame) space-time of target following Hereafter model:

Wherein, ρ is undated parameter.For the first two field picture, its spatial context model is space-time context model.

Step 4, target following

Step 4.1, sets up confidence map

To being obtained space-time context model by former frame i.e. t frame, next frame i.e. t+1 frame is carried out target detection tracking, The most first obtain this frame i.e. context prior model of t+1 frame, such as formula (9), then by space-time context model and present frame The i.e. context prior model of t+1 frame sets up the confidence map based on image block of t+1 frame, such as following formula:

Step 4.2, follows the tracks of target position

What confidence map calculated is the target probit that occurs in each cluster centre, so i.e. recognizing at confidence map maximum probability For being the position at target place, it may be assumed that

Step 4.3, yardstick updates

The profile of target, size are always in change, so also to update therewith to improve robustness scale parameter σ, More new formula is as follows:

Wherein,Represent target location, σ_t+1Represent the scale parameter after updating.

Step 5, robot motion's decision-making

The hardware platform of this method is the robot moving platform carrying kinect, the camera collection on kinect use Scene graph in target detection.For making the target of following of robot energy continuous-stable, use intelligence based on fuzzy control rule Speed governing algorithm, controls the left and right wheel speed of robot.According to robot motion model (see accompanying drawing 2), robot is with linear velocity v During traveling, its left and right wheels speed can be respectively calculated as follows:

Wherein, K steering gain, 2d is robot two-wheeled spacing.

Step 5.1, determines that membership function, fuzzy set carry out obfuscation

Linear function can quickly adjust bigger man-machine distance and man-machine range rate, and the change of shaped form function is smooth, Be conducive to the stationarity controlled.When man-machine distance and man-machine range rate are bigger, use triangular membership；When man-machine When distance is in safety range, use Gaussian membership function.

The effect of obfuscation is that the precise volume of input is changed into obfuscation amount, takes X_rFuzzy subset, in its domain All it is divided into 5 set: " recently (VN) ", " nearly (N) ", " normal (ZE) ", " remote (F) ", " farthest (VF) "；Take v_pxFuzzy son Collection, is all divided into 5 set in its domain: " negative big (NB) ", " negative little (NS) ", " normally (ZE) ", " the least (PS) ", " just Greatly (PB) "；Take the fuzzy subset of ν, its domain be all divided into 5 set: " the least (VL) ", " little (L) ", " in (M) ", " big (H) ", " very big (VH) "；Take Y_r、v_pxFuzzy subset, its domain is all divided into 5 set: " negative big (NB) ", " negative Little (NS) ", " normal (ZE) ", " the least (PS) ", " honest (PB) "；Take the fuzzy subset of K, its domain is all divided into 5 Set: " the least (VL) ", " little (L) ", " in (M) ", " big (H) ", " very big (VH) ".Effective opinion of parameter is obtained by test Territory: X_r∈ [0,3], v_px∈ [-1,1], v ∈ [0,200], Y_r∈ [-1,1], v_py∈ [-1,1], K ∈ [0,3].

Step 5.2, sets up and controls rule

R_1i: if Q₁=A_i and Q₂=B_i, then v=C_i；

R_2i: if Q₃=D_i and Q₄=E_i, then K=F_i。

R_1iOn the basis of linear velocity fuzzy Control rule, Q₁Represent man-machine vertical dimension linguistic variable, Q₂Represent people Machine vertical dimension rate of change linguistic variable.Q₃Represent man-machine horizontal range linguistic variable, Q₄Represent man-machine horizontal range rate of change language Speech variable, ν and K represents datum line speed and turning gain linguistic variable respectively.Their Linguistic Value mould in corresponding domain Stick with paste subset and be respectively A_i、B_i、C_i、D_i、E_i、F_i。

Table 1 datum line velocity ambiguity controls rule

Table 2 is turned Gain Fuzzy control rule table

Robot linear velocity is adjusted according to datum line Fuzzy controller, when man-machine vertical dimension is more than safe distance, For quickly following target, system will increase robot motion's linear velocity；When man-machine vertical dimension is less than safe distance, reduce linear speed Spend with ensure man-machine between safe distance；When man-machine vertical dimension is too small, robot stop motion is in case man-machine collision.According to Turning Gain Fuzzy controller adjusts steering gain, and when man-machine horizontal range is excessive, steering gain increases, and radius of turn subtracts Little, robot traveling process quickly adjusts and turns to ensure that target is in central region position.Rule is as shown in Table 1 and Table 2.

Step 5.3 ambiguity solution

Through logical judgment, utilize centroid method defuzzification.The stability of the Fuzzy control system for being described by rule Problem, can analyze its stability according to fuzzy set theory by relational matrix.

It is complete to one algorithm cycle of this step, after obtaining the target location of t+1 frame, repeat the above steps, more The spatial context model of new t+1 frame and space-time context model, prepare for the next frame i.e. target update of t+2 frame, And carry out robot control according to the result followed the tracks of, robot gather the next frame i.e. scene image of t+2 frame and proceed Aforesaid operations, constantly circulation realizes real-time target following.

Accompanying drawing explanation

Fig. 1 is the flow chart of method involved in the present invention；

Fig. 2 robot motion model；

Detailed description of the invention

Below in conjunction with the accompanying drawings patent of the present invention is further elaborated.

Step 1, scene image is split

Use simple linear Iterative Clustering (SLIC) image is split, utilize the color of image, texture and away from From similarity feature, pixel is clustered, form the image block that the obvious compactedness in border is good.

Step 1.1, obtains local entropy

Step 1.2, clusters pixel

Step 1.3, updates cluster centre

Step 2, image block based on rarefaction representation screens

In order to preferably tackle situation about blocking and improve real-time, image block is carried out rarefaction representation, to different figures Shape block gives different weights, filters out neighbour's block of object block as context area.

Step 2.1, sets up the sparse equation of image block, and composes weight

Step 2.2, solves sparse equation

After obtaining the rarefaction representation of image block, sparse equation is solved, obtain sparse coefficient and weight.

Step 3, sets up space-time context model

Set up space-time context model based on image block obtained above, prepare for target following below.

Step 3.1, sets up the contextual feature of target area cluster centre

Target area contextual feature is obtained in the case of known to target area.

Step 3.2, sets up the context prior model of image block

Then current context prior model is obtained.

Step 3.3, it is thus achieved that spatial context model based on image block

By target location and context priori target, under Bayesian frame, obtain spatial context model.

Step 3.4, space-time updating context

By the weighting of spatial context more newly obtained space-time context model, the target following for next frame is prepared.

Step 4, target following

The space-time context model obtained by previous frame obtains target location confidence with the context prior model of present frame Figure, is considered target location at confidence map maximum probability.After obtaining target location, the spatial context model to frame in the ban is carried out Updating, and then obtain new space-time context model, the target following continuing as next frame is prepared.

Step 4.1, sets up confidence map

Step 4.2, follows the tracks of target position

Step 4.3, yardstick updates

For adapting to target shape in motor process, target area done yardstick and updates by the change of size.

Step 5, robot motion's decision-making

Result according to target location obtained above carries out the motion decision-making of robot, uses the controlling party of fuzzy control Method.

Step 5.1, determines that membership function, fuzzy set carry out obfuscation

Step 5.2, sets up and controls rule

Step 5.3 ambiguity solution.

Claims (2)

1. based on multiple features space-time context robot target identification and motion decision method, it is characterised in that:

The technical scheme that this method uses is as follows:

First, carry out scene image segmentation, use simple linear Iterative Clustering, utilize the color of image, texture and distance Information carries out cluster and forms image block pixel, then gives different weights to each image block and screens, obtains image block Rarefaction representation equation, the most just complete segmentation and the screening of image；The image block obtained with segmentation below after picture segmentation Setting up space-time context model for ultimate unit, in the case of known to present frame target location, i.e. the first frame is by being manually specified, First obtaining the context prior model of image block, be then based under Bayesian frame by target location confidence map and image block is upper The hereafter spatial context model to current frame image of prior model, then carries out spatial context more newly obtained next frame figure The space-time context model of picture, and carry out the renewal of scale parameter；After obtaining the space-time context model of next frame, under treatment Obtained the confidence map of target location during one frame by the context prior model of its image block and space-time context model, confidence map is general Rate maximum is i.e. considered the center of target, and the tracking of target the most just completes；Obtain behind target location again to its space-time Context model is updated, and circulates the purpose reaching real-time tracking target with this；Finally after following the tracks of the position obtaining target, Position according to target carries out the motion decision-making of robot, it is achieved the robot recognition and tracking to target；Specifically include following step Rapid:

Step 1, scene image is split

Utilize the color of image, texture and distance similarity feature that pixel is clustered, use the three-dimensional face of Lab color space Color information and positional information, introduce local entropy and represent the feature of pixel, and formation compactedness is good, border laminating degree high, shape rule Image block then, it is achieved the cluster to pixel；

Step 1.1, obtains local entropy

Local entropy h_iBy formula (1) approximate representation:

$h_i=-\underset{1}{\overset{256}{\Σ}}{p}_i\log p_i---\left(1\right)$

Wherein p_iRepresent that current pixel i accounts for the probability of the total number of local pixel；

Step 1.2, clusters pixel

Set i from initial cluster center C_k=[a_k,b_k,h_k,x_k,y_k] start sampling, for reduce sub pixel misspecification can Energy property and edge noise interference, it is the neighborhood of 3x3 that cluster centre is limited to minimum gradient amount, and each pixel is gathered with arest neighbors Class centre distance D_iCluster, such as formula (2):

$D_i=\mathrm{\μ}\sqrt{{(a_k-a_i)}^2+{(b_k-b_i)}^2}+\mathrm{\μ}|h_k-h_i|+(1-2\mathrm{\μ})\sqrt{{(x_k-x_i)}^2+{(y_k-y_i)}^2}---\left(2\right)$

Wherein a_k,b_k, for the Lab color space [L of pixel k_k,a_k,b_k] color component, h_iFor the local entropy required by formula (1), [x_i,y_i] it is the transverse and longitudinal coordinate of pixel i, μ is weight empirical parameter, takes μ=0.4 here；

Step 1.3, updates cluster centre

Once pixel cluster is to closest to center pixel, then update cluster centre Φ_kReplace putting down of all pixels in cluster areas All vectors:

${\mathrm{\Φ}}_k=\frac{1}{N}\underset{i\∈Z_l}{\Σ}\left[\begin{array}{c}{C}_i\\ S_i\end{array}\right]---\left(3\right)$

Wherein, Z_iRepresent with Φ_kCentered by cluster areas, the number of pixels that N is comprised in representing region, C_iRepresent in cluster The heart, S_iFor two-dimensional spatial location；

Step 2, image block based on rarefaction representation screens

Introduce rarefaction representation and each piecemeal is given different weight, and as target context after arest neighbors region is screened Region unit, it is possible to preferably tackle the situation of target occlusion and improve real-time；

Step 2.1, sets up the sparse equation of image block, and composes weight

For the n-th piecemeal y_n, the rarefaction representation equation of Weight is:

$\underset{W_n,S_n}{\arg min}\left(\right||y_n-D_n{S}_n||+\mathrm{\η}|\left|A_n\right|{|}_2^2+{\mathrm{\μW}}_n\left|\right|A_n-\stackrel{\‾}{A}\left|{|}_2^2\right)---\left(4\right)$

Wherein, cluster centre distance D_nCalculated by formula (2), S_nFor two-dimensional spatial location；A_nRepresent the sparse coefficient of n-th piece to Amount, W_nIt is sparse weight coefficient,Being the average of different piecemeal sparse coefficient in context area, η, μ are positive number constant, use In the impact on rarefaction representation of the specification similarity constraint；

Step 2.2, solves sparse equation

Solve sparse equation and be A_nAnd W_nIt is an optimization problem, solves by iterative method；Main thought assumes that W_nFor Fixed value, updates A_nValue, then with the A calculated_nValue is fixed value, then solves W_n, constantly repeat, until A_nAnd W_nConverge to Local Minimum or reach iterations stop value；Assume weight W of all piecemeals₁,W₂,…,W_mIt is it is known that can by formula (4) Sparse coefficient:

$A_n=M_n+{\mathrm{\μW}}_n{P}_{n}Q\underset{l=1}{\overset{m}{\Σ}}\left(W_l{A}_l\right)/\underset{l=1}{\overset{m}{\Σ}}{W}_l---\left(5\right)$

Wherein, P_n=(D^TD+Λ(ηW_n))^-1,M_n=P_nD_n ^Ty_n,D= Φ_kΛ is unit diagonal matrix；Piecemeal y is obtained by above formula_nSparse coefficient vector A_nAfter, weight is obtained by following formula:

$\underset{W_n}{\arg min}\underset{n=1}{\overset{m}{\Σ}}\left({\mathrm{\μW}}_n\right||A_i-\stackrel{\‾}{A}|{|}^2+{\mathrm{LW}}_n\ln W_n)---\left(6\right)$

Can obtain:

$W_n=e^{(-1-\mathrm{\μ}||A_i-\stackrel{\‾}{A}||/L)}---\left(7\right)$

Wherein, L is Lagrange multiplier,

Iterations limits and is set to t_max, use sparse coefficient A of solution by iterative method the n-th piecemeal_nWith weight W_nStep permissible It is summarized as follows:

Input: (1) D_i,(2)y_i

1) initialize, W_n=1, n=1,2 ..., m

2)

3) A is calculated by formula (11)_n

4) W is calculated by formula (13)_n

5)end while

Output: (1) A_n, (2) W_n

Step 3, sets up space-time context model

To complete the segmentation of image and rarefaction screening in above-mentioned steps, obtained with picture block as ultimate unit is sparse Image, will be that the first frame is directly specified below in the case of known present frame i.e. t frame i.e. target location, set up based on image The context model of block, carries out target following for next frame i.e. t+1 frame and prepares；

Step 3.1, sets up the contextual feature of target area cluster centre

Assume that the context area of target is divided into M_cIndividual block, uses R_nD () represents d of target context region in n-th frame Image block, its cluster centre position is set to CR (d), and visual signature is used insteadRepresent, the contextual feature of cluster centre Definition is as follows:

$x^{ctx}=\{ctx\left(CR\right(d\left)\right)=\left(f_n^d\right(x_n),CR(d\left)\right),d=1,...,M_c\}---\left(8\right)$

Step 3.2, sets up the context prior model of image block

By the visual signature of image block is weighted, set up context prior model based on image block:

$P\left(ctx\right(CR\left(d\right)\left)\right|O)=f_n^d\left(x_n\right)w_{\mathrm{\σ}}\left(CR\right(d)-x_n)W_n---\left(9\right)$

Wherein,w_σFor the weighting function of vision noticing mechanism, this parameter is based on image block Cluster centre determine to the distance of target's center position, the image block that the current location of distance objective is the nearest then gives bigger Weight, its contextual information is more important to the prediction of next frame target location；

Step 3.3, it is thus achieved that spatial context model based on image block

Now to obtain the present frame i.e. target location confidence map of t frame and context prior model, Bayesian frame can obtain The spatial context model of present frame, for improving the calculating time, carries out FFT computing acceleration, the then sky based on image block of t frame Between context model be:

$H_n^{ctx}\left(R_n\right(d\left)\right)=F^{-1}\left(\frac{F\left(C_R\right(x_n\left)\right)}{F\left(f_n^d\right(x_n\left)w_{\mathrm{\σ}}\right(CR\left(d\right)-x_n\left)\right)}\right)---\left(10\right)$

Wherein F^-1() represents Fourier inversion, and F () represents Fourier transformation；

Step 3.4, space-time updating context

It is weighted adding up to spatial context, obtains carrying out for next frame (t+1 frame) the space-time context of target following Model:

$H_{n+1}^{stc}\left(R_{n+1}\right(d\left)\right)=(1-\mathrm{\ρ})H_n^{stc}\left(R_n\right(d\left)\right)+{\mathrm{\ρH}}_n^{ctx}\left(R_n\right(d\left)\right)---\left(11\right)$

Wherein, ρ is undated parameter；For the first two field picture, its spatial context model is space-time context model；

Step 4, target following

Step 4.1, sets up confidence map

To being obtained space-time context model by former frame i.e. t frame, next frame i.e. t+1 frame is carried out target detection tracking, first First obtain this frame i.e. context prior model of t+1 frame, such as formula (9), then by space-time context model and present frame i.e. t The context prior model of+1 frame sets up the confidence map based on image block of t+1 frame, such as following formula:

Step 4.2, follows the tracks of target position

What confidence map calculated is the target probit that occurs in each cluster centre, so being i.e. considered at confidence map maximum probability The position at target place, it may be assumed that

$x_{n+1}^{*}=\underset{x\∈{\mathrm{\Ω}}_{ctx}\left({x_n}^{*}\right)}{\arg }\max C_R\left(x_{n+1}\right)---\left(13\right)$

Step 4.3, yardstick updates

The profile of target, size are always in change, so also to update therewith to improve robustness scale parameter σ, update Formula is as follows:

Wherein,Represent target location, σ_t+1Represent the scale parameter after updating；

Step 5, robot motion's decision-making

The hardware platform of this method is the robot moving platform carrying kinect, by the camera collection on kinect for mesh The scene graph of mark detection；For making the target of following of robot energy continuous-stable, use intelligent speed-regulating based on fuzzy control rule Algorithm, controls the left and right wheel speed of robot；According to robot motion model, when robot advances with linear velocity v, its left side Right wheel speed can be respectively calculated as follows:

$\begin{array}{c}{v}_l=v(1-2{\mathrm{dKY}}_r/(X_r^2+Y_r^2\left)\right)\\ v_r=v(1+2{\mathrm{dKY}}_r/(X_r^2+Y_r^2\left)\right)\end{array}---\left(16\right)$

Wherein, K steering gain, 2d is robot two-wheeled spacing；

Step 5.1, determines that membership function, fuzzy set carry out obfuscation

Linear function can quickly adjust bigger man-machine distance and man-machine range rate, and the change of shaped form function is smooth, favorably In the stationarity controlled；When man-machine distance and man-machine range rate are bigger, use triangular membership；When man-machine distance Time in safety range, use Gaussian membership function；

The effect of obfuscation is that the precise volume of input is changed into obfuscation amount, takes X_rFuzzy subset, all divide in its domain It is 5 set: " recently (VN) ", " nearly (N) ", " normal (ZE) ", " remote (F) ", " farthest (VF) "；Take v_pxFuzzy subset, Its domain is all divided into 5 set: " negative big (NB) ", " negative little (NS) ", " normal (ZE) ", " the least (PS) ", " honest (PB)”；Take the fuzzy subset of ν, its domain is all divided into 5 set: " the least (VL) ", " little (L) ", " in (M) ", " big (H) ", " very big (VH) "；Take Y_r、v_pxFuzzy subset, its domain is all divided into 5 set: " negative big (NB) ", " negative little (NS) ", " normal (ZE) ", " the least (PS) ", " honest (PB) "；Take the fuzzy subset of K, its domain is all divided into 5 collection Close: " the least (VL) ", " little (L) ", " in (M) ", " big (H) ", " very big (VH) "；Effective domain of parameter is obtained by test: X_r∈ [0,3], v_px∈ [-1,1], v ∈ [0,200], Y_r∈ [-1,1], v_py∈ [-1,1], K ∈ [0,3]；

Step 5.2, sets up and controls rule

R_1i: if Q₁=A_i and Q₂=B_i, then v=C_i；

R_2i: if Q₃=D_i and Q₄=E_i, then K=F_i；

R_1iOn the basis of linear velocity fuzzy Control rule, Q₁Represent man-machine vertical dimension linguistic variable, Q₂Represent man-machine vertically Range rate linguistic variable；Q₃Represent man-machine horizontal range linguistic variable, Q₄Represent that man-machine horizontal range rate of change language becomes Amount, ν and K represents datum line speed and turning gain linguistic variable respectively；Their Linguistic Value fuzzy son in corresponding domain Collection is respectively A_i、B_i、C_i、D_i、E_i、F_i；

Table 1 datum line velocity ambiguity controls rule

Table 2 is turned Gain Fuzzy control rule table

Robot linear velocity is adjusted, when man-machine vertical dimension is more than safe distance, for soon according to datum line Fuzzy controller Speed follows target, and system will increase robot motion's linear velocity；When man-machine vertical dimension is less than safe distance, reduce linear velocity with Ensure man-machine between safe distance；When man-machine vertical dimension is too small, robot stop motion is in case man-machine collision；According to turning Gain Fuzzy controller adjusts steering gain, and when man-machine horizontal range is excessive, steering gain increases, and radius of turn reduces, machine Device people quickly adjusts during advancing and turns to ensure that target is in central region position；Rule is as shown in Table 1 and Table 2；

Step 5.3 ambiguity solution

Through logical judgment, utilize centroid method defuzzification；The stability problem of the Fuzzy control system for being described by rule, Its stability can be analyzed by relational matrix according to fuzzy set theory；

It is complete to one algorithm cycle of this step, after obtaining the target location of t+1 frame, repeat the above steps, update t The spatial context model of+1 frame and space-time context model, prepare for the next frame i.e. target update of t+2 frame, and according to The result followed the tracks of carries out robot control, robot gather the next frame i.e. scene image of t+2 frame and proceed above-mentioned behaviour Making, constantly circulation realizes real-time target following.

The most according to claim 1 based on multiple features space-time context robot target identification and motion decision method, its It is characterised by:

Step 1, scene image is split

Use simple linear Iterative Clustering that image is split, utilize the color of image, texture and distance similarity special Levy and pixel is clustered, form the image block that the obvious compactedness in border is good；

Step 1.1, obtains local entropy

Step 1.2, clusters pixel

Step 1.3, updates cluster centre

Step 2, image block based on rarefaction representation screens

In order to preferably tackle situation about blocking and improve real-time, image block is carried out rarefaction representation, to different graph blocks Give different weights, filter out neighbour's block of object block as context area；

Step 2.1, sets up the sparse equation of image block, and composes weight

Step 2.2, solves sparse equation

After obtaining the rarefaction representation of image block, sparse equation is solved, obtain sparse coefficient and weight；

Step 3, sets up space-time context model

Set up space-time context model based on image block obtained above, prepare for target following below；

Step 3.1, sets up the contextual feature of target area cluster centre

Target area contextual feature is obtained in the case of known to target area；

Step 3.2, sets up the context prior model of image block

Then current context prior model is obtained；

Step 3.3, it is thus achieved that spatial context model based on image block

By target location and context priori target, under Bayesian frame, obtain spatial context model；

Step 3.4, space-time updating context

By the weighting of spatial context more newly obtained space-time context model, the target following for next frame is prepared；

Step 4, target following

The space-time context model obtained by previous frame obtains target location confidence map with the context prior model of present frame, puts It is considered target location at letter figure maximum probability；After obtaining target location, the spatial context model to frame in the ban is updated, And then obtaining new space-time context model, the target following continuing as next frame is prepared；

Step 4.1, sets up confidence map

Step 4.2, follows the tracks of target position

Step 4.3, yardstick updates

For adapting to target shape in motor process, target area done yardstick and updates by the change of size；

Step 5, robot motion's decision-making

Result according to target location obtained above carries out the motion decision-making of robot, uses the control method of fuzzy control；

Step 5.1, determines that membership function, fuzzy set carry out obfuscation

Step 5.2, sets up and controls rule

Step 5.3 ambiguity solution.