CN105718896A - Intelligent robot with target recognition function - Google Patents

Abstract

The invention discloses an intelligent robot with a target recognition function. The intelligent robot comprises an intelligent robot and a monitoring device arranged on the intelligent robot, the monitoring device comprises a pre-processing module, a detection tracking module, and a recognition output module, the pre-processing module includes three sub-modules: an image conversion module, an image filtering module, and an image enhancement module, and the detection tracking module includes three sub-modules: a construction module, a loss discrimination module, and an updating module. According to the intelligent robot, the video image technology is applied to the intelligent robot, malicious damage behaviors can be effectively monitored and recorded, and the intelligent robot is advantaged by good timeliness, accurate positioning, high adaptive capability, complete reservation of image details, and high robustness.

Description

A kind of intelligent robot with target recognition function

Technical field

The present invention relates to field in intelligent robotics, be specifically related to a kind of intelligent robot with target recognition function.

Background technology

Intelligent robot possesses panoramic internal information sensor and external information sensor, such as vision, audition, sense of touch, olfactory sensation.Except having sensor, it also has effector, as the means acting on surrounding.Here it is muscles, or claiming motor synchronizing motor, they make hands, foot, long-snouted, feeler etc. move up.Thus will also realize that, intelligent robot at least to possess three key elements: feels key element, kinematicchain element and thinking key element.

As can be seen here, except the organoleptic requirements of the functional requirement of intelligent robot own, it is as a kind of important expensive device, and the safety of intelligent robot is particularly important, it is necessary to can prevent and monitor malicious sabotage behavior.

Summary of the invention

For the problems referred to above, the present invention provides a kind of intelligent robot with target recognition function.

The purpose of the present invention realizes by the following technical solutions:

A kind of intelligent robot with target recognition function, including intelligent robot and the monitoring device being arranged on intelligent robot, monitoring device for carrying out video image monitoring to the activity near intelligent robot, and monitoring device includes pretreatment module, detecting and tracking module, identifies output module；

(1) pretreatment module, for the image received is carried out pretreatment, specifically includes image transformant module, image filtering submodule and image enhaucament submodule:

Image transformant module, for coloured image is converted into gray level image:

$\begin{array}{c}H(x,y)=\frac{\max \left(R\right(x,y),G(x,y),B(x,y\left)\right)+\min \left(R\right(x,y),G(x,y),B(x,y\left)\right)}{2}2\left(\max \right(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right))-\\ \min \left(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right)\right))\end{array}$

Wherein, (x, y), (x, y), (x, (x, y) the intensity red green blue value at place, (x y) represents coordinate (x, y) grey scale pixel value at place to H to B to G to R y) to represent pixel respectively；Image is sized to m × n；

Image filtering submodule, for gray level image is filtered:

Adopt Wiener filtering to carry out after first-level filtering removes, define svlm image, be designated as M_svlm(x, y), being specifically defined formula is: M_svlm(x, y)=a₁J₁(x, y)+a₂J₂(x, y)+a₃J₃(x, y)+a₄J₄(x, y), wherein a₁、a₂、a₃、a₄For variable weight, $a_i=\frac{\mathrm{Ji}}{J1+J2+J3+J4},i=\mathrm{1,2,3,4};$ (x, y) for the image after filtered for J；

Image enhaucament submodule:

When $\left|128-m\right|>\sqrt[3]{\left|\mathrm{\ω}-50\right|}$ Time, $L(x,y)=255\×{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\ψ}(x,y)},$ Wherein, (x, y) for enhanced gray value for L；(x y) is the gamma correction coefficient including local message, now to ψα be range for 0 to 1 variable element,

When $\left|128-m\right|\≤\sqrt[3]{\left|\mathrm{\ω}-50\right|}$ And during ω > 50, $L(x,y)=255\×{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\ψ}(x,y)\×(1-\frac{\mathrm{\ω}-50}{{\mathrm{\ω}}^2})},$ Wherein ψ (x, y)=ψ_α(M_svlm(x, y)),m_HIt is the average of the gray value all pixels higher than 128, m in image_LIt is the average of the gray value all pixels lower than 128, and now m=min (m_H, m_L), when α value is known, calculates 256 ψ correction coefficients as look-up table, be designated asWherein i is index value, utilizes M_svlm(x, gray value y) is as index, according to ψ (x, y)=ψ_α(M_svlm(x, y)) quickly obtain each pixel in image gamma correction coefficient ψ (x, y)；For template correction factor；

(2) detecting and tracking module, specifically includes structure submodule, loses differentiation submodule and update submodule:

Build submodule, for the structure of visual dictionary:

Obtain the position and yardstick of following the tracks of target at initial frame, choosing positive and negative sample training tracker about, result will be followed the tracks of as training set X={x₁, x₂... x_N}^T；And the every width target image in training set is extracted the SIFT feature of 128 dimensionsWherein S_tThe number of SIFT feature in t width target image in expression training set；After following the tracks of N frame, by clustering algorithm, these features are divided into K bunch, the center constitutive characteristic word of each bunch, it is designated asThe feature total amount that can extractWherein K < < F_N, andAfter visual dictionary builds, every width training image is expressed as the form of feature bag, for representing the frequency that in visual dictionary, feature word occurs, with rectangular histogram h (x_t) represent, h (x_t) obtain in the following manner: by a width training image X_tIn each feature f_s ^(t)Projecting to visual dictionary, the feature word the shortest with projector distance represents this feature, after all Projection Characters, adds up the frequency of occurrences of each feature word, and normalization obtains training image X_tFeature histogram h (x_t)；

Lose and differentiate submodule, for differentiating that the loss of target is whether:

When a new two field picture arrives, from K histogram, randomly select K histogram of Z < and Z=4, form the new sub-rectangular histogram h being sized to Z^(z)(x_t), sub histogrammic number is up toIndividual；Calculate candidate target region son histogrammic similarity Φ corresponding to certain target area in training set_{t_z},Wherein t=1,2 ..., N, z=1,2 ..., N_s, then calculate overall similarity Φ_t=1-∏_z(1-Φ_{t_z})；Similarity Φ=max{ Φ of candidate target region and target_t, t} represents, then track rejection judges that formula is: $u=sign\left(\mathrm{\&Phi;}\right)=\left\{\begin{array}{cc}1& \mathrm{\&Phi;}\&GreaterEqual;gs\\ 0& \mathrm{\&Phi;}<gs\end{array}\right.,$ Wherein gs be manually set sentence mistake threshold values；As u=1, target is by tenacious tracking, as u=0, and track rejection；

When track rejection, define affine Transform Model: $\left(\begin{array}{c}{x}_t\\ y_t\end{array}\right)=\left(\begin{array}{cc}s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\\ -s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\end{array}\right)\left(\begin{array}{c}{x}_{t-1}\\ y_{t-1}\end{array}\right)+{\mathrm{\&mu;}}_2\left(\begin{array}{c}e\\ f\end{array}\right),$ Wherein (x_t, y_t) and (x_t-1, y_t-1) the respectively position coordinates of certain SITF characteristic point and the position coordinates of Corresponding matching characteristic point in previous frame target in present frame target, both are known quantity；S is scale coefficient, and θ is coefficient of rotary, and e and f represents translation coefficient, ${\mathrm{\&mu;}}_1=\left\{\begin{array}{cc}1-\frac{\left|T-T_0\right|}{1000T_0}& T\&GreaterEqual;T_0\\ 1+\frac{\left|T-T_0\right|}{1000T_0}& T<T_0\end{array}\right.$ For temperature rotation correction coefficient, ${\mathrm{\&mu;}}_2=\left\{\begin{array}{cc}\sqrt{1-\frac{\left|T-T_0\right|}{1000T_0}}& T\&GreaterEqual;T_0\\ \sqrt{1+\frac{\left|T-T_0\right|}{1000T_0}}& T<T_0\end{array}\right.$ Correction factor, μ is translated for temperature₁And μ₂For revising because the image rotation that causes of ambient temperature deviation and translation error, T₀For the standard temperature being manually set, being set to 20 degree, T is monitored the temperature value obtained in real time by temperature sensor；Adopt Ransac algorithm for estimating to ask for the parameter of affine Transform Model, under new yardstick s and coefficient of rotary θ, finally gather positive negative sample, update grader；

Update submodule, for the renewal of visual dictionary:

After every two field picture obtains target location, the result of calculation according to affine transformation parameter, collect all SIFT feature points meeting result parameterAfter F=3 frame, it is thus achieved that new feature point setWherein S_t-FRepresent the total characteristic obtained from F two field picture to count；Utilize following formula that new and old characteristic point re-starts K cluster: WhereinRepresenting new visual dictionary, the size of visual dictionary remains unchanged；It is forgetting factor, it was shown that proportion shared by old dictionary,More little, the judgement of track rejection is contributed more many by new feature, takes

(3) output module is identified, identification and output for image: utilize track algorithm to obtain target area in image sequence to be identified, target area is mapped to the subspace that known training data is formed, calculate the distance between target area and training data in subspace, obtain similarity measurement, judge target classification, and export recognition result.

Preferably, adopting Wiener filtering to carry out after first-level filtering removes, now image information also includes the noise of remnants, adopts following two-stage filter to carry out secondary filtering:

$J(x,y)={\mathrm{\&Sigma;}}_{i=-m/2}^{m/2}{\mathrm{\&Sigma;}}_{j=-n/2}^{n/2}H(x,y)P_g(x+i,y+j)$

Wherein, J (x, y) be after filtering after image；P_g(x+i, y+j) represents the function that yardstick is m × n and P_g(x+i, y+j)=q × exp (-(x²+y²)/ω), wherein q is by the coefficient of function normalization, it may be assumed that ∫ ∫ q × exp (-(x²+y²)/ω) dxdy=1.

This intelligent robot have the beneficial effect that at image pre-processing phase, the image strengthened can according to the size adaptation adjustment of template, improve reinforced effects, and can automatically revise at the Rule of judgment when different templates size, and consider visual custom and human eye to non-linear relation with colouring intensity of the perceptibility of different color；M × N number of power exponent computing is reduced to 256, improves computational efficiency；At target detection and tracking phase, the error that different temperatures causes the rotation of image and translation to cause can be eliminated, improve discrimination, image detail after treatment becomes apparent from, and amount of calculation is greatly reduced relative to traditional method, can effectively adapt to target scale change, and can accurately judge whether target loses, can by detection tenacious tracking again after target comes back to visual field.Additionally, this intelligent robot has, real-time is good, the advantage of accurate positioning and strong robustness, and achieves good effect in quickly having the target detection blocked and tracking.

Accompanying drawing explanation

The invention will be further described to utilize accompanying drawing, but the embodiment in accompanying drawing does not constitute any limitation of the invention, for those of ordinary skill in the art, under the premise not paying creative work, it is also possible to obtain other accompanying drawing according to the following drawings.

Fig. 1 is the structured flowchart of a kind of intelligent robot with target recognition function；

Fig. 2 is the outside schematic diagram of a kind of intelligent robot with target recognition function.

Detailed description of the invention

The invention will be further described with the following Examples.

Embodiment 1: as shown in Figure 1-2, a kind of intelligent robot with target recognition function, including intelligent robot 5 and the monitoring device 4 being arranged on intelligent robot 5, monitoring device 4 for carrying out video image monitoring to the activity near intelligent robot, and monitoring device 4 includes pretreatment module 1, detecting and tracking module 2, identifies output module 3.

(1) pretreatment module 1, for the image received is carried out pretreatment, specifically includes image transformant module 11, image filtering submodule 12 and image enhaucament submodule 13:

Image transformant module 11, for coloured image is converted into gray level image:

$\begin{array}{c}H(x,y)=\frac{\max \left(R\right(x,y),G(x,y),B(x,y\left)\right)+\min \left(R\right(x,y),G(x,y),B(x,y\left)\right)}{2}2\left(\max \right(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right))-\\ \min \left(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right)\right))\end{array}$

Wherein, (x, y), (x, y), (x, (x, y) the intensity red green blue value at place, (x y) represents coordinate (x, y) grey scale pixel value at place to H to B to G to R y) to represent pixel respectively；Image is sized to m × n；

Image filtering submodule 12, for gray level image is filtered:

Adopt Wiener filtering to carry out after first-level filtering removes, define svlm image, be designated as M_svlm(x, y), being specifically defined formula is: M_svlm(x, y)=a₁J₁(x, y)+a₂J₂(x, y)+a₃J₃(x, y)+a₄J₄(x, y), wherein a₁、a₂、a₃、a₄For variable weight, $a_i=\frac{\mathrm{Ji}}{J1+J2+J3+J4},i=\mathrm{1,2,3,4};$ (x, y) for the image after filtered for J；

Image enhaucament submodule 13:

When $\left|128-m\right|>\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ Time, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)},$ Wherein, (x, y) for enhanced gray value for L；(x y) is the gamma correction coefficient including local message, now to ψα be range for 0 to 1 variable element,

When $\left|128-m\right|\&le;\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ And during ω > 50, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)\&times;(1-\frac{\mathrm{\&omega;}-50}{{\mathrm{\&omega;}}^2})},$ Wherein ψ (x, y)=ψ_α(M_svlm(x, y)),m_HIt is the average of the gray value all pixels higher than 128, m in image_LIt is the average of the gray value all pixels lower than 128, and now m=min (m_H, m_L), when α value is known, calculates 256 ψ correction coefficients as look-up table, be designated asWherein i is index value, utilizes M_svlm(x, gray value y) is as index, according to ψ (x, y)=ψ_α(M_svlm(x, y)) quickly obtain each pixel in image gamma correction coefficient ψ (x, y)；For template correction factor；

(2) detecting and tracking module 2, specifically includes structure submodule 21, loses differentiation submodule 22 and update submodule 23:

Build submodule 21, for the structure of visual dictionary:

Obtain the position and yardstick of following the tracks of target at initial frame, choosing positive and negative sample training tracker about, result will be followed the tracks of as training set X={x₁, x₂... x_N}^T；And the every width target image in training set is extracted the SIFT feature of 128 dimensionsWherein S_tThe number of SIFT feature in t width target image in expression training set；After following the tracks of N frame, by clustering algorithm, these features are divided into K bunch, the center constitutive characteristic word of each bunch, it is designated asThe feature total amount that can extractWherein K < < F_N, andAfter visual dictionary builds, every width training image is expressed as the form of feature bag, for representing the frequency that in visual dictionary, feature word occurs, with rectangular histogram h (x_t) represent, h (x_t) obtain in the following manner: by a width training image X_tIn each feature f_s ^(t)Projecting to visual dictionary, the feature word the shortest with projector distance represents this feature, after all Projection Characters, adds up the frequency of occurrences of each feature word, and normalization obtains training image X_tFeature histogram h (x_t)；

Lose and differentiate submodule 22, for differentiating that the loss of target is whether:

When a new two field picture arrives, from K histogram, randomly select K histogram of Z < and Z=4, form the new sub-rectangular histogram h being sized to Z^(z)(x_t), sub histogrammic number is up toIndividual；Calculate candidate target region son histogrammic similarity Φ corresponding to certain target area in training set_{t_z},Wherein t=1,2 ..., N, z=1,2 ..., N_s, then calculate overall similarity Φ_t=1-∏_z(1-Φ_{t_z})；Similarity Φ=max{ Φ of candidate target region and target_t, t} represents, then track rejection judges that formula is: $u=sign\left(\mathrm{\&Phi;}\right)=\left\{\begin{array}{cc}1& \mathrm{\&Phi;}\&GreaterEqual;gs\\ 0& \mathrm{\&Phi;}<gs\end{array}\right.,$ Wherein gs be manually set sentence mistake threshold values；As u=1, target is by tenacious tracking, as u=0, and track rejection；

When track rejection, define affine Transform Model: $\left(\begin{array}{c}{x}_t\\ y_t\end{array}\right)=\left(\begin{array}{cc}s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\\ -s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\end{array}\right)\left(\begin{array}{c}{x}_{t-1}\\ y_{t-1}\end{array}\right)+{\mathrm{\&mu;}}_2\left(\begin{array}{c}e\\ f\end{array}\right),$ Wherein (x_t, y_t) and (x_t-1, y_t-1) the respectively position coordinates of certain SITF characteristic point and the position coordinates of Corresponding matching characteristic point in previous frame target in present frame target, both are known quantity；S is scale coefficient, and θ is coefficient of rotary, and e and f represents translation coefficient, ${\mathrm{\&mu;}}_1=\left\{\begin{array}{cc}1-\frac{\left|T-T_0\right|}{1000T_0}& T\&GreaterEqual;T_0\\ 1+\frac{\left|T-T_0\right|}{1000T_0}& T<T_0\end{array}\right.$ For temperature rotation correction coefficient, ${\mathrm{\&mu;}}_2=\left\{\begin{array}{cc}\sqrt{1-\frac{\left|T-T_0\right|}{1000T_0}}& T\&GreaterEqual;T_0\\ \sqrt{1+\frac{\left|T-T_0\right|}{1000T_0}}& T<T_0\end{array}\right.$ Correction factor, μ is translated for temperature₁And μ₂For revising because the image rotation that causes of ambient temperature deviation and translation error, T₀For the standard temperature being manually set, being set to 20 degree, T is monitored the temperature value obtained in real time by temperature sensor；Adopt Ransac algorithm for estimating to ask for the parameter of affine Transform Model, under new yardstick s and coefficient of rotary θ, finally gather positive negative sample, update grader；

Update submodule 23, for the renewal of visual dictionary:

After every two field picture obtains target location, the result of calculation according to affine transformation parameter, collect all SIFT feature points meeting result parameterAfter F=3 frame, it is thus achieved that new feature point setWherein S_t-FRepresent the total characteristic obtained from F two field picture to count；Utilize following formula that new and old characteristic point re-starts K cluster: WhereinRepresenting new visual dictionary, the size of visual dictionary remains unchanged；It is forgetting factor, it was shown that proportion shared by old dictionary,More little, the judgement of track rejection is contributed more many by new feature, takes

(3) output module is identified, identification and output for image: utilize track algorithm to obtain target area in image sequence to be identified, target area is mapped to the subspace that known training data is formed, calculate the distance between target area and training data in subspace, obtain similarity measurement, judge target classification, and export recognition result.

Preferably, adopting Wiener filtering to carry out after first-level filtering removes, now image information also includes the noise of remnants, adopts following two-stage filter to carry out secondary filtering:

$J(x,y)={\mathrm{\&Sigma;}}_{i=-m/2}^{m/2}{\mathrm{\&Sigma;}}_{j=-n/2}^{n/2}H(x,y)P_g(x+i,y+j)$

Wherein, J (x, y) be after filtering after image；P_g(x+i, y+j) represents the function that yardstick is m × n and P_g(x+i, y+j)=q × exp (-(x²+y²)/ω), wherein q is by the coefficient of function normalization, it may be assumed that ∫ ∫ q × exp (-(x²+y²)/ω) dxdy=1.

The intelligent robot of this embodiment, at image pre-processing phase, the image strengthened can according to the size adaptation adjustment of template, improve reinforced effects, and can automatically revise at the Rule of judgment when different templates size, and consider visual custom and human eye to non-linear relation with colouring intensity of the perceptibility of different color；Take full advantage of local feature and the global characteristics of image, there is adaptivity, it is possible to suppress excessively to strengthen, the image enhancement effects obtained under complex illumination environment is obvious；M × N number of power exponent computing is reduced to 256, improves computational efficiency, Z=4, F=3,Calculating average frame per second is 15FPS, and amount of calculation is less than the dictionary algorithm of same type；At target detection and tracking phase, the error that different temperatures causes the rotation of image and translation to cause can be eliminated, improve discrimination, image detail after treatment becomes apparent from, and amount of calculation is greatly reduced relative to traditional method, it is possible to effectively adapt to target scale change, and can accurately judge whether target loses, can again be detected and tenacious tracking after target comes back to visual field, until remaining to tenacious tracking target after 110 frames.Additionally, this intelligent robot has, real-time is good, the advantage of accurate positioning and strong robustness, and has good effect in quickly having the target detection blocked and tracking, achieves beyond thought effect.

Embodiment 2: as shown in Figure 1-2, a kind of intelligent robot with target recognition function, including intelligent robot 5 and the monitoring device 4 being arranged on intelligent robot 5, monitoring device 4 for carrying out video image monitoring to the activity near intelligent robot 5, and monitoring device 4 includes pretreatment module 1, detecting and tracking module 2, identifies output module 3.

(1) pretreatment module 1, for the image received is carried out pretreatment, specifically includes image transformant module 11, image filtering submodule 12 and image enhaucament submodule 13:

Image transformant module 11, for coloured image is converted into gray level image:

$\begin{array}{c}H(x,y)=\frac{\max \left(R\right(x,y),G(x,y),B(x,y\left)\right)+\min \left(R\right(x,y),G(x,y),B(x,y\left)\right)}{2}2\left(\max \right(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right))-\\ \min \left(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right)\right))\end{array}$

Wherein, (x, y), (x, y), (x, (x, y) the intensity red green blue value at place, (x y) represents coordinate (x, y) grey scale pixel value at place to H to B to G to R y) to represent pixel respectively；Image is sized to m × n；

Image filtering submodule 12, for gray level image is filtered:

Adopt Wiener filtering to carry out after first-level filtering removes, define svlm image, be designated as M_svlm(x, y), being specifically defined formula is: M_svlm(x, y)=a₁J₁(x, y)+a₂J₂(x, y)+a₃J₃(x, y)+a₄J₄(x, y), wherein a₁、a₂、a₃、a₄For variable weight, $a_i=\frac{\mathrm{Ji}}{J1+J2+J3+J4},i=\mathrm{1,2,3,4};$ (x, y) for the image after filtered for J；

Image enhaucament submodule 13:

When $\left|128-m\right|>\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ Time, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)},$ Wherein, (x, y) for enhanced gray value for L；(x y) is the gamma correction coefficient including local message, now to ψα be range for 0 to 1 variable element,

When $\left|128-m\right|\&le;\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ And during ω > 50, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)\&times;(1-\frac{\mathrm{\&omega;}-50}{{\mathrm{\&omega;}}^2})},$ Wherein ψ (x, y)=ψ_α(M_svlm(x, y)),m_HIt is the average of the gray value all pixels higher than 128, m in image_LIt is the average of the gray value all pixels lower than 128, and now m=min (m_H, m_L), when α value is known, calculates 256 ψ correction coefficients as look-up table, be designated asWherein i is index value, utilizes M_svlm(x, gray value y) is as index, according to ψ (x, y)=ψ_α(M_svlm(x, y)) quickly obtain each pixel in image gamma correction coefficient ψ (x, y)；For template correction factor；

(2) detecting and tracking module 2, specifically includes structure submodule 21, loses differentiation submodule 22 and update submodule 23:

Build submodule 21, for the structure of visual dictionary:

Obtain the position and yardstick of following the tracks of target at initial frame, choosing positive and negative sample training tracker about, result will be followed the tracks of as training set X={x₁, x₂... x_N}^T；And the every width target image in training set is extracted the SIFT feature of 128 dimensionsWherein S_tThe number of SIFT feature in t width target image in expression training set；After following the tracks of N frame, by clustering algorithm, these features are divided into K bunch, the center constitutive characteristic word of each bunch, it is designated asThe feature total amount that can extractWherein K < < F_N, andAfter visual dictionary builds, every width training image is expressed as the form of feature bag, for representing the frequency that in visual dictionary, feature word occurs, with rectangular histogram h (x_t) represent, h (x_t) obtain in the following manner: by a width training image X_tIn each feature f_s ^(t)Projecting to visual dictionary, the feature word the shortest with projector distance represents this feature, after all Projection Characters, adds up the frequency of occurrences of each feature word, and normalization obtains training image X_tFeature histogram h (x_t)；

Lose and differentiate submodule 22, for differentiating that the loss of target is whether:

When a new two field picture arrives, from K histogram, randomly select K histogram of Z < and Z=5, form the new sub-rectangular histogram h being sized to Z^(z)(x_t), sub histogrammic number is up toIndividual；Calculate candidate target region son histogrammic similarity Φ corresponding to certain target area in training set_{t_z},Wherein t=1,2 ..., N, z=1,2 ..., N_s, then calculate overall similarity Φ_t=1-∏_z(1-Φ_{t_z})；Similarity Φ=max{ Φ of candidate target region and target_t, t} represents, then track rejection judges that formula is: $u=sign\left(\mathrm{\&Phi;}\right)=\left\{\begin{array}{cc}1& \mathrm{\&Phi;}\&GreaterEqual;gs\\ 0& \mathrm{\&Phi;}<gs\end{array}\right.,$ Wherein gs be manually set sentence mistake threshold values；As u=1, target is by tenacious tracking, as u=0, and track rejection；

When track rejection, define affine Transform Model: $\left(\begin{array}{c}{x}_t\\ y_t\end{array}\right)=\left(\begin{array}{cc}s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\\ -s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\end{array}\right)\left(\begin{array}{c}{x}_{t-1}\\ y_{t-1}\end{array}\right)+{\mathrm{\&mu;}}_2\left(\begin{array}{c}e\\ f\end{array}\right),$ Wherein (x_t, y_t) and (x_t-1, y_t-1) the respectively position coordinates of certain SITF characteristic point and the position coordinates of Corresponding matching characteristic point in previous frame target in present frame target, both are known quantity；S is scale coefficient, and θ is coefficient of rotary, and e and f represents translation coefficient, ${\mathrm{\&mu;}}_1=\left\{\begin{array}{cc}1-\frac{\left|T-T_0\right|}{1000T_0}& T\&GreaterEqual;T_0\\ 1+\frac{\left|T-T_0\right|}{1000T_0}& T<T_0\end{array}\right.$ For temperature rotation correction coefficient, ${\mathrm{\&mu;}}_2=\left\{\begin{array}{cc}\sqrt{1-\frac{\left|T-T_0\right|}{1000T_0}}& T\&GreaterEqual;T_0\\ \sqrt{1+\frac{\left|T-T_0\right|}{1000T_0}}& T<T_0\end{array}\right.$ Correction factor, μ is translated for temperature₁And μ₂For revising because the image rotation that causes of ambient temperature deviation and translation error, T₀For the standard temperature being manually set, being set to 20 degree, T is monitored the temperature value obtained in real time by temperature sensor；Adopt Ransac algorithm for estimating to ask for the parameter of affine Transform Model, under new yardstick s and coefficient of rotary θ, finally gather positive negative sample, update grader；

Update submodule 23, for the renewal of visual dictionary:

After every two field picture obtains target location, the result of calculation according to affine transformation parameter, collect all SIFT feature points meeting result parameterAfter F=4 frame, it is thus achieved that new feature point setWherein S_t-FRepresent the total characteristic obtained from F two field picture to count；Utilize following formula that new and old characteristic point re-starts K cluster: WhereinRepresenting new visual dictionary, the size of visual dictionary remains unchanged；It is forgetting factor, it was shown that proportion shared by old dictionary,More little, the judgement of track rejection is contributed more many by new feature, takes

(3) output module is identified, identification and output for image: utilize track algorithm to obtain target area in image sequence to be identified, target area is mapped to the subspace that known training data is formed, calculate the distance between target area and training data in subspace, obtain similarity measurement, judge target classification, and export recognition result.

Preferably, adopting Wiener filtering to carry out after first-level filtering removes, now image information also includes the noise of remnants, adopts following two-stage filter to carry out secondary filtering:

$J(x,y)={\mathrm{\&Sigma;}}_{i=-m/2}^{m/2}{\mathrm{\&Sigma;}}_{j=-n/2}^{n/2}H(x,y)P_g(x+i,y+j)$

Wherein, J (x, y) be after filtering after image；P_g(x+i, y+j) represents the function that yardstick is m × n and P_g(x+i, y+j)=q × exp (-(x²+y²)/ω), wherein q is by the coefficient of function normalization, it may be assumed that ∫ ∫ q × exp (-(x²+y²)/ω) dxdy=1.

The intelligent robot of this embodiment, at image pre-processing phase, the image strengthened can according to the size adaptation adjustment of template, improve reinforced effects, and can automatically revise at the Rule of judgment when different templates size, and consider visual custom and human eye to non-linear relation with colouring intensity of the perceptibility of different color；Take full advantage of local feature and the global characteristics of image, there is adaptivity, it is possible to suppress excessively to strengthen, the image enhancement effects obtained under complex illumination environment is obvious；M × N number of power exponent computing is reduced to 256, improves computational efficiency, Z=5, F=4,Calculating average frame per second is 16FPS, and amount of calculation is less than the dictionary algorithm of same type；At target detection and tracking phase, the error that different temperatures causes the rotation of image and translation to cause can be eliminated, improve discrimination, image detail after treatment becomes apparent from, and amount of calculation is greatly reduced relative to traditional method, it is possible to effectively adapt to target scale change, and can accurately judge whether target loses, can again be detected and tenacious tracking after target comes back to visual field, until remaining to tenacious tracking target after 115 frames.Additionally, this intelligent robot has, real-time is good, the advantage of accurate positioning and strong robustness, and has good effect in quickly having the target detection blocked and tracking, achieves beyond thought effect.

Embodiment 3: as shown in Figure 1-2, a kind of intelligent robot with target recognition function, including intelligent robot 5 and the monitoring device 4 being arranged on intelligent robot 5, monitoring device 4 for carrying out video image monitoring to the activity near intelligent robot 5, and monitoring device 4 includes pretreatment module 1, detecting and tracking module 2, identifies output module 3.

(1) pretreatment module 1, for the image received is carried out pretreatment, specifically includes image transformant module 11, image filtering submodule 12 and image enhaucament submodule 13:

Image transformant module 11, for coloured image is converted into gray level image:

$\begin{array}{c}H(x,y)=\frac{\max \left(R\right(x,y),G(x,y),B(x,y\left)\right)+\min \left(R\right(x,y),G(x,y),B(x,y\left)\right)}{2}2\left(\max \right(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right))-\\ \min \left(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right)\right))\end{array}$

Wherein, (x, y), (x, y), (x, (x, y) the intensity red green blue value at place, (x y) represents coordinate (x, y) grey scale pixel value at place to H to B to G to R y) to represent pixel respectively；Image is sized to m × n；

Image filtering submodule 12, for gray level image is filtered:

Adopt Wiener filtering to carry out after first-level filtering removes, define svlm image, be designated as M_svlm(x, y), being specifically defined formula is: M_svlm(x, y)=a₁J₁(x, y)+a₂J₂(x, y)+a₃J₃(x, y)+a₄J₄(x, y), wherein a₁、a₂、a₃、a₄For variable weight, $a_i=\frac{\mathrm{Ji}}{J1+J2+J3+J4},i=\mathrm{1,2,3,4};$ (x, y) for the image after filtered for J；

Image enhaucament submodule 13:

When $\left|128-m\right|>\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ Time, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)},$ Wherein, (x, y) for enhanced gray value for L；(x y) is the gamma correction coefficient including local message, now to ψα be range for 0 to 1 variable element,

When $\left|128-m\right|\&le;\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ And during ω > 50, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)\&times;(1-\frac{\mathrm{\&omega;}-50}{{\mathrm{\&omega;}}^2})},$ Wherein ψ (x, y)=ψ_α(M_svlm(x, y)),m_HIt is the average of the gray value all pixels higher than 128, m in image_LIt is the average of the gray value all pixels lower than 128, and now m=min (m_H, m_L), when α value is known, calculates 256 ψ correction coefficients as look-up table, be designated asWherein i is index value, utilizes M_svlm(x, gray value y) is as index, according to ψ (x, y)=ψ_α(M_svlm(x, y)) quickly obtain each pixel in image gamma correction coefficient ψ (x, y)；For template correction factor；

(2) detecting and tracking module 2, specifically includes structure submodule 21, loses differentiation submodule 22 and update submodule 23:

Build submodule 21, for the structure of visual dictionary:

Obtain the position and yardstick of following the tracks of target at initial frame, choosing positive and negative sample training tracker about, result will be followed the tracks of as training set X={x₁, x₂... x_N}^T；And the every width target image in training set is extracted the SIFT feature of 128 dimensionsWherein S_tThe number of SIFT feature in t width target image in expression training set；After following the tracks of N frame, by clustering algorithm, these features are divided into K bunch, the center constitutive characteristic word of each bunch, it is designated asThe feature total amount that can extractWherein K < < F_N, andAfter visual dictionary builds, every width training image is expressed as the form of feature bag, for representing the frequency that in visual dictionary, feature word occurs, with rectangular histogram h (x_t) represent, h (x_t) obtain in the following manner: by a width training image X_tIn each feature f_s ^(t)Projecting to visual dictionary, the feature word the shortest with projector distance represents this feature, after all Projection Characters, adds up the frequency of occurrences of each feature word, and normalization obtains training image X_tFeature histogram h (x_t)；

Lose and differentiate submodule 22, for differentiating that the loss of target is whether:

When a new two field picture arrives, from K histogram, randomly select K histogram of Z < and Z=6, form the new sub-rectangular histogram h being sized to Z^(z)(x_t), sub histogrammic number is up toIndividual；Calculate candidate target region son histogrammic similarity Φ corresponding to certain target area in training set_{t_z},Wherein t=1,2 ..., N, z=1,2 ..., N_s, then calculate overall similarity Φ_t=1-∏_z(1-Φ_{t_z})；Similarity Φ=max{ Φ of candidate target region and target_t, t} represents, then track rejection judges that formula is: $u=sign\left(\mathrm{\&Phi;}\right)=\left\{\begin{array}{cc}1& \mathrm{\&Phi;}\&GreaterEqual;gs\\ 0& \mathrm{\&Phi;}<gs\end{array}\right.,$ Wherein gs be manually set sentence mistake threshold values；As u=1, target is by tenacious tracking, as u=0, and track rejection；

When track rejection, define affine Transform Model: $\left(\begin{array}{c}{x}_t\\ y_t\end{array}\right)=\left(\begin{array}{cc}s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\\ -s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\end{array}\right)\left(\begin{array}{c}{x}_{t-1}\\ y_{t-1}\end{array}\right)+{\mathrm{\&mu;}}_2\left(\begin{array}{c}e\\ f\end{array}\right),$ Wherein (x_t, y_t) and (x_t-1, y_t-1) the respectively position coordinates of certain SITF characteristic point and the position coordinates of Corresponding matching characteristic point in previous frame target in present frame target, both are known quantity；S is scale coefficient, and θ is coefficient of rotary, and e and f represents translation coefficient, ${\mathrm{\&mu;}}_1=\left\{\begin{array}{cc}1-\frac{\left|T-T_0\right|}{1000T_0}& T\&GreaterEqual;T_0\\ 1+\frac{\left|T-T_0\right|}{1000T_0}& T<T_0\end{array}\right.$ For temperature rotation correction coefficient, ${\mathrm{\&mu;}}_2=\left\{\begin{array}{cc}\sqrt{1-\frac{\left|T-T_0\right|}{1000T_0}}& T\&GreaterEqual;T_0\\ \sqrt{1+\frac{\left|T-T_0\right|}{1000T_0}}& T<T_0\end{array}\right.$ Correction factor, μ is translated for temperature₁And μ₂For revising because the image rotation that causes of ambient temperature deviation and translation error, T₀For the standard temperature being manually set, being set to 20 degree, T is monitored the temperature value obtained in real time by temperature sensor；Adopt Ransac algorithm for estimating to ask for the parameter of affine Transform Model, under new yardstick s and coefficient of rotary θ, finally gather positive negative sample, update grader；

Update submodule 23, for the renewal of visual dictionary:

After every two field picture obtains target location, the result of calculation according to affine transformation parameter, collect all SIFT feature points meeting result parameterAfter F=5 frame, it is thus achieved that new feature point setWherein S_t-FRepresent the total characteristic obtained from F two field picture to count；Utilize following formula that new and old characteristic point re-starts K cluster: WhereinRepresenting new visual dictionary, the size of visual dictionary remains unchanged；It is forgetting factor, it was shown that proportion shared by old dictionary,More little, the judgement of track rejection is contributed more many by new feature, takes

(3) output module 3 is identified, identification and output for image: utilize track algorithm to obtain target area in image sequence to be identified, target area is mapped to the subspace that known training data is formed, calculate the distance between target area and training data in subspace, obtain similarity measurement, judge target classification, and export recognition result.

Preferably, adopting Wiener filtering to carry out after first-level filtering removes, now image information also includes the noise of remnants, adopts following two-stage filter to carry out secondary filtering:

$J(x,y)={\mathrm{\&Sigma;}}_{i=-m/2}^{m/2}{\mathrm{\&Sigma;}}_{j=-n/2}^{n/2}H(x,y)P_g(x+i,y+j)$

Wherein, J (x, y) be after filtering after image；P_g(x+i, y+j) represents the function that yardstick is m × n and P_g(x+i, y+j)=q × exp (-(x²+y²)/ω), wherein q is by the coefficient of function normalization, it may be assumed that ∫ ∫ q × exp (-(x²+y²)/ω) dxdy=1.

The intelligent robot of this embodiment, at image pre-processing phase, the image strengthened can according to the size adaptation adjustment of template, improve reinforced effects, and can automatically revise at the Rule of judgment when different templates size, and consider visual custom and human eye to non-linear relation with colouring intensity of the perceptibility of different color；Take full advantage of local feature and the global characteristics of image, there is adaptivity, it is possible to suppress excessively to strengthen, the image enhancement effects obtained under complex illumination environment is obvious；M × N number of power exponent computing is reduced to 256, improves computational efficiency, Z=6, F=5,Calculating average frame per second is 17FPS, and amount of calculation is less than the dictionary algorithm of same type；At target detection and tracking phase, the error that different temperatures causes the rotation of image and translation to cause can be eliminated, improve discrimination, image detail after treatment becomes apparent from, and amount of calculation is greatly reduced relative to traditional method, it is possible to effectively adapt to target scale change, and can accurately judge whether target loses, can again be detected and tenacious tracking after target comes back to visual field, until remaining to tenacious tracking target after 120 frames.Additionally, this intelligent robot has, real-time is good, the advantage of accurate positioning and strong robustness, and has good effect in quickly having the target detection blocked and tracking, achieves beyond thought effect.

Embodiment 4: as shown in Figure 1-2, a kind of intelligent robot with target recognition function, including intelligent robot 5 and the monitoring device 4 being arranged on intelligent robot 5, monitoring device 4 for carrying out video image monitoring to the activity near intelligent robot 5, and monitoring device 4 includes pretreatment module 1, detecting and tracking module 2, identifies output module 3.

(1) pretreatment module 1, for the image received is carried out pretreatment, specifically includes image transformant module 11, image filtering submodule 12 and image enhaucament submodule 13:

Image transformant module 11, for coloured image is converted into gray level image:

$\begin{array}{c}H(x,y)=\frac{\max \left(R\right(x,y),G(x,y),B(x,y\left)\right)+\min \left(R\right(x,y),G(x,y),B(x,y\left)\right)}{2}2\left(\max \right(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right))-\\ \min \left(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right)\right))\end{array}$

Wherein, (x, y), (x, y), (x, (x, y) the intensity red green blue value at place, (x y) represents coordinate (x, y) grey scale pixel value at place to H to B to G to R y) to represent pixel respectively；Image is sized to m × n；

Image filtering submodule 12, for gray level image is filtered:

Adopt Wiener filtering to carry out after first-level filtering removes, define svlm image, be designated as M_svlm(x, y), being specifically defined formula is: M_svlm(x, y)=a₁J₁(x, y)+a₂J₂(x, y)+a₃J₃(x, y)+a₄J₄(x, y), wherein a₁、a₂、a₃、a₄For variable weight, $a_i=\frac{\mathrm{Ji}}{J1+J2+J3+J4},i=\mathrm{1,2,3,4};$ (x, y) for the image after filtered for J；

Image enhaucament submodule 13:

When $\left|128-m\right|>\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ Time, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)},$ Wherein, (x, y) for enhanced gray value for L；(x y) is the gamma correction coefficient including local message, now to ψα be range for 0 to 1 variable element,

When $\left|128-m\right|\&le;\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ And during ω > 50, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)\&times;(1-\frac{\mathrm{\&omega;}-50}{{\mathrm{\&omega;}}^2})},$ Wherein ψ (x, y)=ψ_α(M_svlm(x, y)),m_HIt is the average of the gray value all pixels higher than 128, m in image_LIt is the average of the gray value all pixels lower than 128, and now m=min (m_H, m_L), when α value is known, calculates 256 ψ correction coefficients as look-up table, be designated asWherein i is index value, utilizes M_svlm(x, gray value y) is as index, according to ψ (x, y)=ψ_α(M_svlm(x, y)) quickly obtain each pixel in image gamma correction coefficient ψ (x, y)；For template correction factor；

(2) detecting and tracking module 2, specifically includes structure submodule 21, loses differentiation submodule 22 and update submodule 23:

Build submodule 21, for the structure of visual dictionary:

Obtain the position and yardstick of following the tracks of target at initial frame, choosing positive and negative sample training tracker about, result will be followed the tracks of as training set X={x₁, x₂... x_N}^T；And the every width target image in training set is extracted the SIFT feature of 128 dimensionsWherein S_tThe number of SIFT feature in t width target image in expression training set；After following the tracks of N frame, by clustering algorithm, these features are divided into K bunch, the center constitutive characteristic word of each bunch, it is designated asThe feature total amount that can extractWherein K < < F_N, andAfter visual dictionary builds, every width training image is expressed as the form of feature bag, for representing the frequency that in visual dictionary, feature word occurs, with rectangular histogram h (x_t) represent, h (x_t) obtain in the following manner: by a width training image X_tIn each feature f_s ^(t)Projecting to visual dictionary, the feature word the shortest with projector distance represents this feature, after all Projection Characters, adds up the frequency of occurrences of each feature word, and normalization obtains training image X_tFeature histogram h (x_t)；

Lose and differentiate submodule 22, for differentiating that the loss of target is whether:

When a new two field picture arrives, from K histogram, randomly select K histogram of Z < and Z=7, form the new sub-rectangular histogram h being sized to Z^(z)(x_t), sub histogrammic number is up toIndividual；Calculate candidate target region son histogrammic similarity Φ corresponding to certain target area in training set_{t_z},Wherein t=1,2 ..., N, z=1,2 ..., N_s, then calculate overall similarity Φ_t=1-∏_z(1-Φ_{t_z})；Similarity Φ=max{ Φ of candidate target region and target_t, t} represents, then track rejection judges that formula is: $u=sign\left(\mathrm{\&Phi;}\right)=\left\{\begin{array}{cc}1& \mathrm{\&Phi;}\&GreaterEqual;gs\\ 0& \mathrm{\&Phi;}<gs\end{array}\right.,$ Wherein gs be manually set sentence mistake threshold values；As u=1, target is by tenacious tracking, as u=0, and track rejection；

When track rejection, define affine Transform Model: $\left(\begin{array}{c}{x}_t\\ y_t\end{array}\right)=\left(\begin{array}{cc}s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\\ -s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\end{array}\right)\left(\begin{array}{c}{x}_{t-1}\\ y_{t-1}\end{array}\right)+{\mathrm{\&mu;}}_2\left(\begin{array}{c}e\\ f\end{array}\right),$ Wherein (x_t, y_t) and (x_t-1, y_t-1) the respectively position coordinates of certain SITF characteristic point and the position coordinates of Corresponding matching characteristic point in previous frame target in present frame target, both are known quantity；S is scale coefficient, and θ is coefficient of rotary, and e and f represents translation coefficient, ${\mathrm{\&mu;}}_1=\left\{\begin{array}{cc}1-\frac{\left|T-T_0\right|}{1000T_0}& T\&GreaterEqual;T_0\\ 1+\frac{\left|T-T_0\right|}{1000T_0}& T<T_0\end{array}\right.$ For temperature rotation correction coefficient, ${\mathrm{\&mu;}}_2=\left\{\begin{array}{cc}\sqrt{1-\frac{\left|T-T_0\right|}{1000T_0}}& T\&GreaterEqual;T_0\\ \sqrt{1+\frac{\left|T-T_0\right|}{1000T_0}}& T<T_0\end{array}\right.$ Correction factor, μ is translated for temperature₁And μ₂For revising because the image rotation that causes of ambient temperature deviation and translation error, T₀For the standard temperature being manually set, being set to 20 degree, T is monitored the temperature value obtained in real time by temperature sensor；Adopt Ransac algorithm for estimating to ask for the parameter of affine Transform Model, under new yardstick s and coefficient of rotary θ, finally gather positive negative sample, update grader；

Update submodule 23, for the renewal of visual dictionary:

After every two field picture obtains target location, the result of calculation according to affine transformation parameter, collect all SIFT feature points meeting result parameterAfter F=6 frame, it is thus achieved that new feature point setWherein S_t-FRepresent the total characteristic obtained from F two field picture to count；Utilize following formula that new and old characteristic point re-starts K cluster: WhereinRepresenting new visual dictionary, the size of visual dictionary remains unchanged；It is forgetting factor, it was shown that proportion shared by old dictionary,More little, the judgement of track rejection is contributed more many by new feature, takes

(3) output module 3 is identified, identification and output for image: utilize track algorithm to obtain target area in image sequence to be identified, target area is mapped to the subspace that known training data is formed, calculate the distance between target area and training data in subspace, obtain similarity measurement, judge target classification, and export recognition result.

Preferably, adopting Wiener filtering to carry out after first-level filtering removes, now image information also includes the noise of remnants, adopts following two-stage filter to carry out secondary filtering:

$J(x,y)={\mathrm{\&Sigma;}}_{i=-m/2}^{m/2}{\mathrm{\&Sigma;}}_{j=-n/2}^{n/2}H(x,y)P_g(x+i,y+j)$

Wherein, J (x, y) be after filtering after image；P_g(x+i, y+j) represents the function that yardstick is m × n and P_g(x+i, y+j)=q × exp (-(x²+y²)/ω), wherein q is by the coefficient of function normalization, it may be assumed that ∫ ∫ q × exp (-(x²+y²)/ω) dxdy=1.

The intelligent robot of this embodiment, at image pre-processing phase, the image strengthened can according to the size adaptation adjustment of template, improve reinforced effects, and can automatically revise at the Rule of judgment when different templates size, and consider visual custom and human eye to non-linear relation with colouring intensity of the perceptibility of different color；Take full advantage of local feature and the global characteristics of image, there is adaptivity, it is possible to suppress excessively to strengthen, the image enhancement effects obtained under complex illumination environment is obvious；M × N number of power exponent computing being reduced to 256, improves computational efficiency, Z=7, F=6, φ=0.18, calculating average frame per second is 18FPS, and amount of calculation is less than the dictionary algorithm of same type；At target detection and tracking phase, the error that different temperatures causes the rotation of image and translation to cause can be eliminated, improve discrimination, image detail after treatment becomes apparent from, and amount of calculation is greatly reduced relative to traditional method, it is possible to effectively adapt to target scale change, and can accurately judge whether target loses, can again be detected and tenacious tracking after target comes back to visual field, until remaining to tenacious tracking target after 125 frames.Additionally, this intelligent robot has, real-time is good, the advantage of accurate positioning and strong robustness, and has good effect in quickly having the target detection blocked and tracking, achieves beyond thought effect.

Embodiment 5: as shown in Figure 1-2, a kind of intelligent robot with target recognition function, including intelligent robot 5 and the monitoring device 4 being arranged on intelligent robot 5, monitoring device 4 for carrying out video image monitoring to the activity near intelligent robot 5, and monitoring device 4 includes pretreatment module 1, detecting and tracking module 2, identifies output module 3.

(1) pretreatment module 1, for the image received is carried out pretreatment, specifically includes image transformant module 11, image filtering submodule 12 and image enhaucament submodule 13:

Image transformant module 11, for coloured image is converted into gray level image:

$\begin{array}{c}H(x,y)=\frac{\max \left(R\right(x,y),G(x,y),B(x,y\left)\right)+\min \left(R\right(x,y),G(x,y),B(x,y\left)\right)}{2}2\left(\max \right(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right))-\\ \min \left(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right)\right))\end{array}$

Wherein, (x, y), (x, y), (x, (x, y) the intensity red green blue value at place, (x y) represents coordinate (x, y) grey scale pixel value at place to H to B to G to R y) to represent pixel respectively；Image is sized to m × n；

Image filtering submodule 12, for gray level image is filtered:

Adopt Wiener filtering to carry out after first-level filtering removes, define svlm image, be designated as M_svlm(x, y), being specifically defined formula is: M_svlm(x, y)=a₁J₁(x, y)+a₂J₂(x, y)+a₃J₃(x, y)+a₄J₄(x, y), wherein a₁、a₂、a₃、a₄For variable weight, $a_i=\frac{\mathrm{Ji}}{J1+J2+J3+J4},i=\mathrm{1,2,3,4};$ (x, y) for the image after filtered for J；

Image enhaucament submodule 13:

When $\left|128-m\right|>\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ Time, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)},$ Wherein, (x, y) for enhanced gray value for L；(x y) is the gamma correction coefficient including local message, now to ψα be range for 0 to 1 variable element,

When $\left|128-m\right|\&le;\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ And during ω > 50, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)\&times;(1-\frac{\mathrm{\&omega;}-50}{{\mathrm{\&omega;}}^2})},$ Wherein ψ (x, y)=ψ_α(M_svlm(x, y)),m_HIt is the average of the gray value all pixels higher than 128, m in image_LIt is the average of the gray value all pixels lower than 128, and now m=min (m_H, m_L), when α value is known, calculates 256 ψ correction coefficients as look-up table, be designated asWherein i is index value, utilizes M_svlm(x, gray value y) is as index, according to ψ (x, y)=ψ_α(M_svlm(x, y)) quickly obtain each pixel in image gamma correction coefficient ψ (x, y)；For template correction factor；

(2) detecting and tracking module 2, specifically includes structure submodule 21, loses differentiation submodule 22 and update submodule 23:

Build submodule 21, for the structure of visual dictionary:

Obtain the position and yardstick of following the tracks of target at initial frame, choosing positive and negative sample training tracker about, result will be followed the tracks of as training set X={x₁, x₂... x_N}^T；And the every width target image in training set is extracted the SIFT feature of 128 dimensionsWherein S_tThe number of SIFT feature in t width target image in expression training set；After following the tracks of N frame, by clustering algorithm, these features are divided into K bunch, the center constitutive characteristic word of each bunch, it is designated asThe feature total amount that can extractWherein K < < F_N, andAfter visual dictionary builds, every width training image is expressed as the form of feature bag, for representing the frequency that in visual dictionary, feature word occurs, with rectangular histogram h (x_t) represent, h (x_t) obtain in the following manner: by a width training image X_tIn each feature f_s ^(t)Projecting to visual dictionary, the feature word the shortest with projector distance represents this feature, after all Projection Characters, adds up the frequency of occurrences of each feature word, and normalization obtains training image X_tFeature histogram h (x_t)；

Lose and differentiate submodule 22, for differentiating that the loss of target is whether:

When a new two field picture arrives, from K histogram, randomly select K histogram of Z < and Z=8, form the new sub-rectangular histogram h being sized to Z^(z)(x_t), sub histogrammic number is up toIndividual；Calculate candidate target region son histogrammic similarity Φ corresponding to certain target area in training set_{t_z},Wherein t=1,2 ..., N, z=1,2 ..., N_s, then calculate overall similarity Φ_t=1-∏_z(1-Φ_{t_z})；Similarity Φ=max{ Φ of candidate target region and target_t, t} represents, then track rejection judges that formula is: $u=sign\left(\mathrm{\&Phi;}\right)=\left\{\begin{array}{cc}1& \mathrm{\&Phi;}\&GreaterEqual;gs\\ 0& \mathrm{\&Phi;}<gs\end{array}\right.,$ Wherein gs be manually set sentence mistake threshold values；As u=1, target is by tenacious tracking, as u=0, and track rejection；

When track rejection, define affine Transform Model: $\left(\begin{array}{c}{x}_t\\ y_t\end{array}\right)=\left(\begin{array}{cc}s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\\ -s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\end{array}\right)\left(\begin{array}{c}{x}_{t-1}\\ y_{t-1}\end{array}\right)+{\mathrm{\&mu;}}_2\left(\begin{array}{c}e\\ f\end{array}\right),$ Wherein (x_t, y_t) and (x_t-1, y_t-1) the respectively position coordinates of certain SITF characteristic point and the position coordinates of Corresponding matching characteristic point in previous frame target in present frame target, both are known quantity；S is scale coefficient, and θ is coefficient of rotary, and e and f represents translation coefficient, ${\mathrm{\&mu;}}_1=\left\{\begin{array}{cc}1-\frac{\left|T-T_0\right|}{1000T_0}& T\&GreaterEqual;T_0\\ 1+\frac{\left|T-T_0\right|}{1000T_0}& T<T_0\end{array}\right.$ For temperature rotation correction coefficient, ${\mathrm{\&mu;}}_2=\left\{\begin{array}{cc}\sqrt{1-\frac{\left|T-T_0\right|}{1000T_0}}& T\&GreaterEqual;T_0\\ \sqrt{1+\frac{\left|T-T_0\right|}{1000T_0}}& T<T_0\end{array}\right.$ Correction factor, μ is translated for temperature₁And μ₂For revising because the image rotation that causes of ambient temperature deviation and translation error, T₀For the standard temperature being manually set, being set to 20 degree, T is monitored the temperature value obtained in real time by temperature sensor；Adopt Ransac algorithm for estimating to ask for the parameter of affine Transform Model, under new yardstick s and coefficient of rotary θ, finally gather positive negative sample, update grader；

Update submodule 23, for the renewal of visual dictionary:

After every two field picture obtains target location, the result of calculation according to affine transformation parameter, collect all SIFT feature points meeting result parameterAfter F=7 frame, it is thus achieved that new feature point setWherein S_t-FRepresent the total characteristic obtained from F two field picture to count；Utilize following formula that new and old characteristic point re-starts K cluster: WhereinRepresenting new visual dictionary, the size of visual dictionary remains unchanged；It is forgetting factor, it was shown that proportion shared by old dictionary,More little, the judgement of track rejection is contributed more many by new feature, takes

(3) output module 3 is identified, identification and output for image: utilize track algorithm to obtain target area in image sequence to be identified, target area is mapped to the subspace that known training data is formed, calculate the distance between target area and training data in subspace, obtain similarity measurement, judge target classification, and export recognition result.

Preferably, adopting Wiener filtering to carry out after first-level filtering removes, now image information also includes the noise of remnants, adopts following two-stage filter to carry out secondary filtering:

$J(x,y)={\mathrm{\&Sigma;}}_{i=-m/2}^{m/2}{\mathrm{\&Sigma;}}_{j=-n/2}^{n/2}H(x,y)P_g(x+i,y+j)$

Wherein, J (x, y) be after filtering after image；P_g(x+i, y+j) represents the function that yardstick is m × n and P_g(x+i, y+j)=q × exp (-(x²+y²)/ω), wherein q is by the coefficient of function normalization, it may be assumed that ∫ ∫ q × exp (-(x²+y²)/ω) dxdy=1.

The intelligent robot of this embodiment, at image pre-processing phase, the image strengthened can according to the size adaptation adjustment of template, improve reinforced effects, and can automatically revise at the Rule of judgment when different templates size, and consider visual custom and human eye to non-linear relation with colouring intensity of the perceptibility of different color；Take full advantage of local feature and the global characteristics of image, there is adaptivity, it is possible to suppress excessively to strengthen, the image enhancement effects obtained under complex illumination environment is obvious；M × N number of power exponent computing is reduced to 256, improves computational efficiency, Z=8, F=7,Calculating average frame per second is 19FPS, and amount of calculation is less than the dictionary algorithm of same type；At target detection and tracking phase, the error that different temperatures causes the rotation of image and translation to cause can be eliminated, improve discrimination, image detail after treatment becomes apparent from, and amount of calculation is greatly reduced relative to traditional method, it is possible to effectively adapt to target scale change, and can accurately judge whether target loses, can again be detected and tenacious tracking after target comes back to visual field, until remaining to tenacious tracking target after 130 frames.Additionally, this intelligent robot has, real-time is good, the advantage of accurate positioning and strong robustness, and has good effect in quickly having the target detection blocked and tracking, achieves beyond thought effect.

Claims (2)

1. an intelligent robot with target recognition function, including intelligent robot and the monitoring device being arranged on intelligent robot, monitoring device is for carrying out video image monitoring to the activity near intelligent robot, it is characterized in that, monitoring device includes pretreatment module, detecting and tracking module, identifies output module；

(1) pretreatment module, for the image received is carried out pretreatment, specifically includes image transformant module, image filtering submodule and image enhaucament submodule:

Image transformant module, for coloured image is converted into gray level image:

$\begin{array}{c}H(x,y)=\frac{\max \left(R\right(x,y),G(x,y),B(x,y\left)\right)+\min \left(R\right(x,y),G(x,y),B(x,y\left)\right)}{2}+2\left(\max \right(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right))-\\ \min \left(R\left(x,y\right),G\left(x,y\right),B\left(x,y\right)\right))\end{array}$

Wherein, (x, y), (x, y), (x, (x, y) the intensity red green blue value at place, (x y) represents coordinate (x, y) grey scale pixel value at place to H to B to G to R y) to represent pixel respectively；Image is sized to m × n；

Image filtering submodule, for gray level image is filtered:

Adopt Wiener filtering to carry out after first-level filtering removes, define svlm image, be designated as M_svlm(x, y), being specifically defined formula is: M_svlm(x, y)=a₁J₁(x, y)+a₂J₂(x, y)+a₃J₃(x, y)+a₄J₄(x, y), wherein a₁、a₂、a₃、a₄For variable weight,I=1,2,3,4；(x, y) for the image after filtered for J；

Image enhaucament submodule:

WhenTime, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)},$ Wherein, (x, y) for enhanced gray value for L；(x y) is the gamma correction coefficient including local message, now to ψα be range for 0 to 1 variable element,ω is template scale size parameter, and yardstick is more big, and the neighborhood territory pixel information comprised in template is more many, and input picture is through different scale ω_iTemplate, the image J obtained_iThe neighborhood information of different range will be comprised；

When $\left|128-m\right|\&le;\sqrt[3]{\left|\mathrm{\&omega;}-50\right|}$ And during ω > 50, $L(x,y)=255\&times;{\left(\frac{H(x,y)}{255}\right)}^{\mathrm{\&psi;}(x,y)\&times;(1-\frac{\mathrm{\&omega;}-50}{{\mathrm{\&omega;}}^2})},$ Wherein ψ (x, y)=ψ_α(M_svlm(x, y)),m_HIt is the average of the gray value all pixels higher than 128, m in image_LIt is the average of the gray value all pixels lower than 128, and now m=min (m_H, m_L), when α value is known, calculates 256 ψ correction coefficients as look-up table, be designated asWherein i is index value, utilizes M_svlm(x, gray value y) is as index, according to ψ (x, y)=ψ_α(M_svlm(x, y)) quickly obtain each pixel in image gamma correction coefficient ψ (x, y)；For template correction factor；

(2) detecting and tracking module, specifically includes structure submodule, loses differentiation submodule and update submodule:

Build submodule, for the structure of visual dictionary:

Obtain the position and yardstick of following the tracks of target at initial frame, choosing positive and negative sample training tracker about, result will be followed the tracks of as training set X={x₁, x₂... x_N}^T；And the every width target image in training set is extracted the SIFT feature of 128 dimensionsWherein S_tThe number of SIFT feature in t width target image in expression training set；After following the tracks of N frame, by clustering algorithm, these features are divided into K bunch, the center constitutive characteristic word of each bunch, it is designated asThe feature total amount that can extractWherein K < < F_N, andAfter visual dictionary builds, every width training image is expressed as the form of feature bag, for representing the frequency that in visual dictionary, feature word occurs, with rectangular histogram h (x_t) represent, h (x_t) obtain in the following manner: by a width training image X_tIn each feature f_s ^(t)Projecting to visual dictionary, the feature word the shortest with projector distance represents this feature, after all Projection Characters, adds up the frequency of occurrences of each feature word, and normalization obtains training image X_tFeature histogram h (x_t)；

Lose and differentiate submodule, for differentiating that the loss of target is whether:

When a new two field picture arrives, from K histogram, randomly select K histogram of Z < and Z=4, form the new sub-rectangular histogram h being sized to Z^(z)(x_t), sub histogrammic number is up toIndividual；Calculate candidate target region son histogrammic similarity Φ corresponding to certain target area in training set_{t_z},Wherein t=1,2 ..., N, z=1,2 ..., N_s, then calculate overall similarity Φ_t=1-∏_z(1-Φ_{t_z})；Similarity Φ=max{ Φ of candidate target region and target_t, t} represents, then track rejection judges that formula is: $u=sign\left(\mathrm{\&Phi;}\right)=\left\{\begin{array}{cc}1& \mathrm{\&Phi;}\&GreaterEqual;gs\\ 0& \mathrm{\&Phi;}<gs\end{array}\right.,$ Wherein gs be manually set sentence mistake threshold values；As u=1, target is by tenacious tracking, as u=0, and track rejection；When track rejection, define affine Transform Model: $\left(\begin{array}{c}{x}_t\\ y_t\end{array}\right)=\left(\begin{array}{cc}s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\\ -s.sin({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})& s.cos({\mathrm{\&mu;}}_1\&times;\mathrm{\&theta;})\end{array}\right)\left(\begin{array}{c}{x}_{t-1}\\ y_{t-1}\end{array}\right)+{\mathrm{\&mu;}}_2\left(\begin{array}{c}e\\ f\end{array}\right),$ Wherein (x_t, y_t) and (x_t-1, y_t-1) the respectively position coordinates of certain SITF characteristic point and the position coordinates of Corresponding matching characteristic point in previous frame target in present frame target, both are known quantity；S is scale coefficient, and θ is coefficient of rotary, and e and f represents translation coefficient, ${\mathrm{\&mu;}}_1=\left\{\begin{array}{cc}1-\frac{\left|T-T_0\right|}{1000T_0}& T\&GreaterEqual;T_0\\ 1+\frac{\left|T-T_0\right|}{1000T_0}& T<T_0\end{array}\right.$ For temperature rotation correction coefficient, ${\mathrm{\&mu;}}_2=\left\{\begin{array}{cc}\sqrt{1-\frac{\left|T-T_0\right|}{1000T_0}}& T\&GreaterEqual;T_0\\ \sqrt{1+\frac{\left|T-T_0\right|}{1000T_0}}& T<T_0\end{array}\right.$ Correction factor, μ is translated for temperature₁And μ₂For revising because the image rotation that causes of ambient temperature deviation and translation error, T₀For the standard temperature being manually set, being set to 20 degree, T is monitored the temperature value obtained in real time by temperature sensor；Adopt Ransac algorithm for estimating to ask for the parameter of affine Transform Model, under new yardstick s and coefficient of rotary θ, finally gather positive negative sample, update grader；

Update submodule, for the renewal of visual dictionary:

After every two field picture obtains target location, the result of calculation according to affine transformation parameter, collect all SIFT feature points meeting result parameterAfter F=3 frame, it is thus achieved that new feature point setWherein S_t-FRepresent the total characteristic obtained from F two field picture to count；Utilize following formula that new and old characteristic point re-starts K cluster: WhereinRepresenting new visual dictionary, the size of visual dictionary remains unchanged；It is forgetting factor, it was shown that proportion shared by old dictionary,More little, the judgement of track rejection is contributed more many by new feature, takes

(3) output module is identified, identification and output for image: utilize track algorithm to obtain target area in image sequence to be identified, target area is mapped to the subspace that known training data is formed, calculate the distance between target area and training data in subspace, obtain similarity measurement, judge target classification, and export recognition result.

2. a kind of intelligent robot with target recognition function according to claim 1, is characterized in that, adopts Wiener filtering to carry out after first-level filtering removes, and now image information also includes the noise of remnants, adopts following two-stage filter to carry out secondary filtering:

$J(x,y)={\mathrm{\&Sigma;}}_{i=-m/2}^{m/2}{\mathrm{\&Sigma;}}_{j=-n/2}^{n/2}H(x,y)P_g(x+i,y+j)$

Wherein, J (x, y) be after filtering after image；P_g(x+i, y+j) represents the function that yardstick is m × n and P_g(x+i, y+j)=q × exp (-(x²+y²)/ω), wherein q is by the coefficient of function normalization, it may be assumed that ∫ ∫ q × exp (-(x²+y²)/ω) dxdy=1.