CN1604125A - Target image automatic identification and rapid tracking method - Google Patents

Abstract

It is an aim image automatic identifying and rapid tracing method, which comprises the following steps: first to establish aim image valve grey scale grade; second to establish error and correlation index; then to form mode image and spot image characteristics matrix on the base of valve value grey scale and to form correlation functions under the condition of discreteness, such as variance, normalization correlation index, inner product operation in the linear space of all two-dimension matrix; to get the correlation definition according to norm and angle in the Euclidean space starting from the linear space; then to get a judge formula by use of inner product definition to realize the rapid identification of the aim.

Description

Target image is identification and the method for following the tracks of fast automatically

Technical field

What the present invention relates to is the method that a kind of target image that is used for technical field of image processing is handled automatically, specifically is a kind of target image identification and the method for following the tracks of fast automatically.

Background technology

Visual picture signal has occupied crucial status day by day in the mankind's social activities, obtained application more and more widely at a lot of subjects such as traffic, monitoring, biomedicine, remote sensing, military affairs, astronomy, geology and geography and field.Picture signal is the 2D signal on a kind of temporal peacekeeping space.The data volume of picture signal is huge, if be subjected to the restriction of data storage capacity or transmission bandwidth, system need be with compression of images to reduce data volume.Simultaneously, the digital processing operational form complexity of higher-dimension, therefore efficient and the speed to algorithm has proposed very high requirement naturally.In addition, the digitizing of picture signal is expressed and is demonstrated its distinctive diversity, the color that different systems selects different modes to express each pixel.System selects a kind of digitizing expression to finish main work of treatment, also will realize the conversion of several modes simultaneously.In addition, picture signal is again a non-negative two-dimensional random field.Because spatially, general picture signal can not be expressed with any analytical function, can not be by typical signal stack, and be exactly statistical method so the research image mainly adopts.In other words, as a general image system, the definite content of input picture can't be predicted by system, therefore can only adopt the method for extracting statistical characteristic value to come the cognitive map picture, thereby finish the work of coupling.

Find by prior art documents, " flotation froth feature and the state recognition thereof " that Liu Wenli etc. deliver (" Chinese coal " fifth phase in 2003) article proposes the coal slime flotation process control thinking based on Digital Image Processing and recognition technology, introduced by space gray scale correlation matrix method and neighborhood gray scale correlation matrix and extracted the method for foam texture features parameter, and utilized self organizing neural network that the coal slime flotation foam state is discerned.Wherein, space gray scale correlation matrix is (θ=0 ° on the different directions of floatation foam image, 45 °, 90 °, 135 °) construct, in view of the dimension of space gray scale correlation matrix big, article is with " thickness " of foam characteristics texture, " trend " waits the numerical characteristic amount that defines, promptly with energy, entropy and moment of inertia are characteristic parameter, can give expression to coal slime flotation foam visual signature, neighborhood gray scale correlation matrix is generalized to the face neighborhood with the notion of line neighborhood and derives, neighborhood gray scale correlation matrix is when extracting the feature of image, considered in the image gray-scale value of all pixels of (face neighborhood) on a certain picture element 8-neighborhood direction as a whole, neighborhood gray scale correlation matrix has extracted the fineness parameter of describing floatation foam image, the rugosity parameter, entropy parameter, second order is apart from parameter and unevenness parameter.The flotation of coal slime is based in this research, be conceived to the statistical appraisal of whole flotation results, do not establish the discrimination index of tracked target body, the element of setting up correlation matrix simultaneously is mainly based on the fineness degree of object, therefore usable range is obviously received limitation, and operand is bigger.Result of study shows: fineness parameter, rugosity parameter can reflect the feature of floatation foam image well, and entropy parameter, second order are not strong apart from the correlativity that parameter and unevenness parameter and image foam characteristic change.Therefore, the method that the document provided is not suitable for identification and the tracking to intended target.

Summary of the invention

The objective of the invention is to overcome deficiency of the prior art and defective, a kind of target image identification and the method for following the tracks of fast automatically are provided, making it be applicable to the recognition and tracking of target image in a lot of subjects such as traffic, monitoring, biomedicine, remote sensing, military affairs, astronomy, geology and geography and the field, is a kind of very practical and effective image processing techniques.

The present invention is achieved by the following technical solutions, the present invention at first sets up target image threshold value gray level, next sets up sum of errors degree of correlation index, then on the basis of threshold value gray level, make up template image and image scene eigenmatrix, and the cross correlation function under discrete conditions, comprise each auto-variance, normalized correlation coefficient, and the definition of inner product operation in the linear space that all two-dimensional matrixs are formed, from linear space, the definition that obtains the degree of correlation according to the norm and the angle of Euclidean space, utilize definition of inner product to obtain the criterion formula of a practicality again, final realization is to the quick identification and the tracking of target.

Below the inventive method is further described:

At first, set up the threshold value gray level,, be divided into black pixel and than the threshold value gray level dark pixel brighter and white two groups than threshold value gray level for brightness value that can split image.Desirable threshold value is divided target: the black and white composition in the image equates substantially, the sharpness of border of image, main body can be differentiated substantially.Be expressed as a vertical separator bar at the gray level histogram upper threshold value, all gray levels on the separator bar left side will become black, and the right side will become white.On the one hand, the area about separator bar should make equates, to guarantee to have identical black and white pixel; On the other hand, suppose that the gray level probability distribution can approach with two Gaussian distribution, one of them represents the main body prospect, another represents unwanted background object, threshold value should be chosen in the interface point of peak valley, can distinguish that to guarantee the two-value sharpness of border foreground area and background area are correctly cut apart.In concrete recognition and tracking process, the present invention is fast searching peak valley interface point at first, existing side by side, soon it is defined as the threshold value of gray level, and then with the foundation of provincial characteristics as image segmentation, come separation prospect and background area, and, realize the good unchangeability of convergent-divergent, translation, rotation and brightness of characteristics of image with the statistical characteristics of second-order moment around mean as coupling.

Secondly, set up sum of errors degree of correlation index, whether mate, when how many similar in other words degree has, at first can use the image feature amount notion, relatively the feature of template image and on-the-spot a certain parts of images at the judge image.If both differ too big, just think not to be same target, if error is in allowed limits, think that then the possibility that they are correlated with is very big.It is the ratio of the difference absolute value of image scene feature and template image eigenwert to the template characteristic value that the present invention defines error.

Then, on the basis of threshold value gray level, quote quality, axle distribution, barycenter and moment of inertia notion in the physics, be based upon normalization on the statistical significance second order mix central moment, set up template image matrix A and image scene matrix B, and the cross correlation function under discrete conditions, comprise the definition of each auto-variance, normalized correlation coefficient and inner product operation in the linear space that all two-dimensional matrixs are formed.From linear space, the definition that obtains the degree of correlation according to the norm and the angle notion of Euclidean space.

At last, utilize definition of inner product can obtain the criterion formula of a practicality.

Angle is more little in linear space, and similarity is big more, and image correlativity is obvious more, and image is similar more, if the degree of correlation reaches 100%, A and B just can go out by mutual linear list so.By seeking the maximal value of the degree of correlation, can find place best on the image scene with the template image linearity, also can find the place of both simple crosscorrelation maximums.Therefore can think that the acquire a certain degree image of (as 95%) of the degree of correlation promptly is same width of cloth image.

The present invention both had been conceived to the whole statistical appraisal of differentiating, the discrimination index of clear and definite again tracked target body, simultaneously realize that less than the superelevation arithmetic speed of 40ms automatic identification and tracking to intended target, device technique performance index reach the correct recognition rata more than 96% and can discern yardstick less than 1m with total processing cycle.

Embodiment

Content below in conjunction with the inventive method provides following examples:

Present embodiment with high-speed photoelectric coupler as the high-speed image sampling device, hi-vision acquisition rate can reach 60 frames/s, image acquisition device is installed on and accepts the orientation controlling and driving on the univesal wheelwork, be connected with the embedded hardware system through image card, use algorithm programming of the present invention, according to being identified the situation that target departs from the center, visual field, the There Axis Turn Table of ordering about image acquisition device by control algolithm realizes being identified the tracking of target.

Concrete enforcement is as follows:

Move with high-speed motion carrier (F-Zero can reach 190km/h), carrier of the present invention (comprising: straight line with the different modes of advancing with the non-uniform velocity of 80～180km/h from distance objective 10km, curve, go up a slope, the mode of advancing such as descending) to target approaches, in the high-speed motion, the present invention is at first to template image (target image of capturing in distance objective 5km place in advance) and image scene fast searching gray scale peak valley interface point, existing side by side, soon it is defined as the threshold value of gray level, and then with the foundation of provincial characteristics as image segmentation, come separation prospect and background area, and, realize the convergent-divergent of characteristics of image with the statistical characteristics of second-order moment around mean as coupling, translation, rotation and the good unchangeability of brightness; Setting up sum of errors degree of correlation index, be the error definition with the difference absolute value of presence feature and template characteristic value to the ratio of template characteristic value, and definite error amount＜2.5% is a permissible error; Similar degree＞95% with template image matrix and image scene matrix is a degree of correlation index, as passing judgment on the foundation whether image mates; Set up template image matrix A and image scene matrix B, and the cross correlation function under discrete conditions, determine degree of correlation definition according to Euclidean space:

$\mathrm{sim}=\cos \∠(A,B)=\frac{A\·B}{\left|\right|A\left|\right|\left|\right|B\left|\right|}=\frac{A\·B}{\sqrt{A\·A}\sqrt{B\·B}}$

Utilize definition of inner product to obtain the Practical Criteria formula:

$\mathrm{sim}=\frac{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}A(x,y)\·B(x,y)}{\sqrt{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{A}^2(x,y)}\sqrt{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{B}^2(x,y)}}$

Implementation result:

Measured result: search the peak valley interface point, determine the threshold value of gray level, consuming time less than 8ms; Separate prospect and background area with provincial characteristics, set up the coupling statistical characteristics, consuming time less than 20ms; It is consuming time less than 10ms to set up normalization second order mixing central moment; Calculate that criterion is consuming time less than 2ms.

Through (with the non-uniform velocity of 80～180km/h and the different modes of the advancing) running test fast of the height rolling ground in radius 5km scope repeatedly, the result proves: under the situation about stopping at random be subjected to mirror under complex background before, to straight line sighting distance 10km Target Recognition accuracy rate up to more than 96%, the total execution cycle of total system recognition and tracking reaches below the 40ms, recognition and tracking speed very obvious soon.

Claims (5)

1, a kind of target image is identification and the method for following the tracks of fast automatically, it is characterized in that, at first set up target image threshold value gray level, next sets up sum of errors degree of correlation index, then on the basis of threshold value gray level, make up template image and image scene eigenmatrix, and the cross correlation function under discrete conditions, comprise each auto-variance, normalized correlation coefficient, and the definition of inner product operation in the linear space that all two-dimensional matrixs are formed, from linear space, the definition that obtains the degree of correlation according to the norm and the angle of Euclidean space, utilize definition of inner product to obtain the criterion formula of a practicality again, final realization is to the quick identification and the tracking of target.

2, target image according to claim 1 identification and the method for following the tracks of fast automatically, it is characterized in that, described threshold value gray level, with the peak valley interface point is threshold value, be divided into black pixel and than the threshold value gray level dark pixel brighter and white two groups than threshold value gray level, fast searching peak valley interface point, existing side by side, soon it is defined as the threshold value of gray level, and then with the foundation of provincial characteristics as image segmentation, come separation prospect and background area, and, realize that convergent-divergent, translation, rotation and the brightness of characteristics of image is good with the statistical characteristics of second-order moment around mean as coupling.

3, target image according to claim 1 identification and the method for following the tracks of fast automatically, it is characterized in that, described template image and image scene feature, the definition error is the ratio of the difference absolute value of image scene feature and template image eigenwert to the template characteristic value, relatively the feature of template image and on-the-spot a certain parts of images if error is too big, is just thought not to be same target, if error in allowed limits, think that then the possibility that they are correlated with is very big;

4, according to claim 1 or the identification and the method for following the tracks of fast automatically of 2 described target images, it is characterized in that, described threshold value gray level, on the basis of threshold value gray level, quote the quality in the physics, axle distributes, barycenter and moment of inertia notion, be based upon normalization on the statistical significance second order mix central moment, set up template image matrix A and image scene matrix B, and the cross correlation function under discrete conditions, comprise each auto-variance, normalized correlation coefficient, and the definition of inner product operation in the linear space that all two-dimensional matrixs are formed, from linear space, establish degree of correlation formula according to the norm and the angle notion of Euclidean space;

5, target image according to claim 1 is identification and the method for following the tracks of fast automatically, it is characterized in that, described definition of inner product, utilize definition of inner product to release practical criterion formula from degree of correlation formula, angle is more little in linear space, similarity is big more, image correlativity is obvious more, image is similar more, if the degree of correlation reaches 100%, A and B just can go out by mutual linear list so, by seeking the maximal value of the degree of correlation, can find place best on the image scene, or find the place of both simple crosscorrelation maximums, so the image that the degree of correlation acquires a certain degree promptly is same width of cloth image with the template image linearity.