Method of object detection

FIELD: automatics.

SUBSTANCE: high efficiency method of object detection is developed. It includes preliminary detection of object search zones, image scanning with the help of scanning window within the limits of the search zones, the size of the scanning window corresponds to the size of an object's image, quantity determination of amplitude intervals N by the square of the scanning window, taking into account the most significant amplitude intervals and choosing the coordinates with maximum value N as the coordinates of the object. The claimed invention increase probability of correct object detection within lower contrasts range of the object and higher contrast range of background.

EFFECT: exclusion of false object detection, which is determined by contrast jump in brightness of stretched background, as well as background which is alike to the object and has smaller size than one of the scanning window.

4 cl, 2 dwg

Method of object detection refers to the automation and technical Cybernetics and can be used in the development of systems for the automatic analysis and classification of images.

When solving problems automatically detect there is a problem of selection of the image object on the natural motley background (earth, air), when the object is known only approximate dimensions. The main feature that distinguishes the image of the object from the background, is the shape formed by the difference of brightness of the image signal, therefore, when analyzing Phono target environment (FCO) usually consider two types of the most unfavorable background noise:

contrasting the differences of brightness of the image, due to border longest backgrounds (for example, when analyzing air FCO - edge cloud, in the analysis of terrestrial FCO - differential brightness "meadow-arable land");

- background image, signature signs coinciding with the image of the object.

The solution of the problem of automatic detection and recognition of moving objects is simplified when using the infrared wavelength range. In this range, first, increases the signal-to-background due to heating of individual parts of objects (e.g., motors), and secondly, there is a natural heat transfer elements of the natural background, different heated, AFL is stvie different orientation relative to the Sun and other heat sources.

There is a method of object detection by filtering medium frequencies (PSC), consisting in the creation of a frame image in the IR range, converting the image into a set of signals representing a rectangular matrix of signals, sampling of signals, one-dimensional digital filtering of high frequencies in rows and columns with parameters consistent with the size of the object, the histogram of brightness of the image, the choice of the quantization threshold according to the histogram of brightness and coding image segmentation, where the detected object is selected, the image is above the threshold level (Levshin V.L. Spatial filtering in optical systems locations. - M.: Owls. radio, 1971). This method is the upper limit of spatial frequencies to improve the signal-to-noise ratio, on the one hand, and the suppression of the contrast of low-frequency components of the lengthy background of formations to improve the signal/background, on the other hand. The advantage of this method is the simplicity and the efficiency of detecting the contrast of low-contrast objects on the background, such as torch aircraft object in the analysis of air FCO and mobile equipment for the winter uniform background in the analysis of terrestrial FCO. However, found the situation less contrasting objects or objects at a more contrasting backgrounds, for example, less heated plating aircraft carrier in the background, the edges of the clouds or the car is turned off and cooled down the engine, the contrast of the background image at the output of the filter medium frequencies may exceed the contrast of the object image, which leads to false detection.

The closest to the technical nature of the proposed solution is a method of object detection, which consists in creating a picture frame in the IR range, converting the image into a set of signals representing a rectangular matrix of signals, sampling the signals, the selection signals as image extended platforms, thermal signature which is determined by the distribution of point sources of infrared energy, the scan image of the field of view of the scanning window with dimensions consistent with the dimensions of the image of the object, determining the area of the scan window the number of N signals with different amplitudes and the choice of coordinates window coordinates of the scan with the maximum value of N (RF Patent No. 2219564, G01S 17/06, publ. 20.12.03). This solution is chosen for the prototype.

In accordance with the prototype within the scanning window with center at the point (i, j) determine the number of amplitude intervals N(i, j) with different amplitudes, which are estimated by considering a non-zero value is s frequency falling timing of the image signals in k-e amplitude intervals for discrete elements of the matrix signals U_in(i, j)belonging to the image of the selected object

where K is the number of quantization levels of the image signals U_in(i, j)

and the coordinates of the objectdefined as the coordinates of the maximum amount of amplitude intervals N(i, j) with different amplitudes

The advantage of this solution is that the detected object, including those with less contrast than the contrast of the background, by taking into account the amplitude range of the brightness distribution of the image detail of the object within the scanning window, the size of which is coordinated with the size of the object.

The disadvantage of this solution is the false detection of the object, first, the contrasting differences of brightness extensive backgrounds, secondly, contrasting objectpooling backgrounds smaller than the scan window.

The task of the invention is to increase the probability of correct detection of the object at lower contrasts of the object and at higher contrasts the background.

The problem is solved due to the fact that there is a way to detect the object side is present in the creation of the frame image in the IR range, converting the image into a set of signals representing a matrix of signals, sampling the signals, the allocation of extended signals by scanning the field of view of the image scan window with dimensions consistent with the dimensions of the image of the object, determining the area of the scan window the number of amplitude intervals N, the coordinates of the object coordinates with a maximum value N.

From the prototype of the inventive method is characterized by the fact that pre-determine the area of search objects, and then in determining the amount of amplitude intervals N consider only the most significant amplitude intervals.

In addition, the search area of interest can be determined by two-dimensional even filtering by rows and columns, consistent with the size of the object in rows and columns, respectively, and two-dimensional odd filter in rows and columns, consistent with the size of the object in rows and columns, respectively, the formation of the criterion function in accordance with the results of the even and odd filter and the centers of the areas of search objectpooling background of such centres of the areas of search, in which the merit function is above a threshold level, and the even and odd filter can be implemented using what W as the pulse characteristics of even and odd the first harmonics of the number of Walsh, implemented recursively in accordance with the patent of the Russian Federation No. 41938. To account for the most significant amplitude intervals when estimating N in the area of the first scanning window can be formed of a brightness histogram by counting the frequency of contact of the brightness values of the image in the appropriate amplitude intervals, and then considered only the amplitude of the intervals where the values of the histogram exceeds a threshold level.

Achievable technical result is the removal of false detections of the object, due to the contrast differences of the brightness of the lengthy background and objectpooling background with dimensions smaller than the size of the scanning window.

The technical result is achieved, first, by pre-determining the areas of search object, which is objectmodule image that prevents false detection of the object contrast differences of the brightness image backgrounds; second, due to the fact that when evaluating the amplitude of the range of values in the scan window, the search does not count all the amplitude intervals of brightness of the object, but only the most significant, which eliminates errors in which the object can be selected background, for example, smaller than the object.

The proposed method is implemented using the algorithm, reflecting the th stages of processing digitized video signal for making decisions about the detected object.

1. Define the search area objectpooling image.

To do this:

1.1 Formed the two-even one-dimensional filtering of the images in rows and columns

As the pulse characteristics are used first even harmonic Walsh, consistent with the size of the object along the lines of l_o.iand columns of l_o.j

To reduce computational cost of implementation of the convolution algorithm (4) with pulse characteristics (5), (6) you can use a digital transversal filter (Patent RF №41938), in accordance with which the convolution (4) can be represented

1.2. Formed the two odd-dimensional filtering of the images in rows and columns as in (7) with a pulse characteristics corresponding to the first odd harmonics Walsh

1.3. Formed criterial function

where σ - stabilizing element whose value is comparable to the RMS value of the noise at the output of the filter mid-frequency.

1.4. Determine the coordinates of the centers (i, j) areas of search objects with subjectpagename images corresponding to the to slovio

where R_thenthe threshold value chosen a priori on the basis of diversity of background.

2. Determine the coordinates of the object. To do this:

2.1. Within the scanning window with center at the point (i, j) defines the most significant amplitude intervals

where g_pthe threshold value that determines the level of the most significant amplitude intervals.

2.2. Within the scanning window as in (1) is the number of most significant amplitude intervals

2.3. Determine the coordinates of the objectin areas of the search object corresponding to the coordinates of the center objectpooling image (10)

Figure 1 presents an example of FCO, including the object with the coordinate of the i_o, a steep drop brightness extensive background with the coordinate of the i_toand objectpool interference with the coordinate of the i_n(for the most simple case, when the video FCO U_in(i) one-dimensional).

In figure 1, and illustrates the original FCO (solid line).

Figure 1, b shows the results of processing FCO using the algorithm (4)...(12) (solid thick line - processing filter (7) with an even pulse response, a continuous thin line - processing filter () with odd impulse response, dotted criterion function (9)).

Figure 2 presents the distribution density of the signal g(U) for 3 amplitude intervals,,corresponding spatial intervals with coordinates at the points of i_o, i_k, i_n(solid line).

In the proposed decision to exclude from the consideration of extended variations of brightness of the selected area object search by finding objectpooling image, which is achieved by formation results not only even (7) (Fig 1, b - thick line)and odd filter mid-frequency (8) (Fig 1, b - thin line) and the subsequent formation of the criterion function (9) (Fig 1, b - dashed line). Use the filtering medium frequencies with odd impulse response leads to resonant underline fronts extensive backgrounds and, as a consequence, the minimum response criterion function (9) (Fig 1, b - dashed line). On the contrary, in the centre objectpooling backgrounds odd results filtering medium frequencies (8) lead to the minimum module response and, consequently, to a sharp increase in the criterion function (9). Therefore, further thresholding (10) allows to exclude from consideration the differences of brightness extensive backgrounds, regardless of what x contrast.

To avoid errors associated with the determination of the object objectmodel interference, are excluded from consideration objectmodule backgrounds similar to how the proposed prototype. The greatest number of amplitude intervals (1) corresponds to the difference of brightness of the background, which is excluded from consideration (Fig 1, b - dashed line in the neighborhood of i_k). Therefore, in FCO (figure 1, solid line), the object will be detected correctly. However, possible FCO, in which, for example, objectbody background has a size smaller than the size of the scanning window object (figure 1, a - dotted line). This leads to an increase in the area of determining the density g(U), which corresponds to the increase in the number of amplitude intervals of N (2 - dashed line) and in the case of the prototype leads to a false detection of a background object. To avoid this error in the proposed method takes into account not all amplitude intervals, as in the prototype (2), but only the most significant (11), i.e. above the threshold level g_p. When g_p=0, the method of identifying the significant amplitude of the intervals in the proposed solution coincides with the method of accounting amplitude intervals of the prototype. Different selection algorithms g_p. For example, selecting the threshold values of g_pas a proportion of the maximum values of the distribution densities in suitable the x amplitude intervals ,,(2) accounting for significant amplitude intervals leads to smaller values of the number of significant intervals of noise N(i, j) and limited intervals,.

Thus, the proposed solution in contrast to the known analogs and prototypes allows you to get additional positive effect. In the above example, FCO of the proposed method are excluded from consideration not only the differences of brightness long, but objectpooling backgrounds.

So, if the analogue and the prototype of the detection signal is dependent on the ratio of contrast of the object and the background, the shape of the background, the proposed solution allows to detect low contrast objects with almost any contrasts and shapes in the background, which provides a very large gain in signal-to-background. The main limitation of this method is the low signal-to-noise ratio. However, as shown by the results of mathematical and scaled-down simulation, taking into account the real model image objects, backgrounds, and noise, the proposed solution is superior to known analogs and the prototype in relation signal/noise ratio of about 3-10 times (depending on the type of FTO).

1. Method of object detection, which increation of the frame image in the IR range, converting the image into a set of signals representing a matrix of signals, sampling the signals, the allocation of extended signals by scanning the field of view of the image scan window with dimensions consistent with the dimensions of the image of the object, determining the area of the scan window the number of amplitude intervals N, the coordinates of the object coordinates with a maximum value of N, wherein the pre-defined area of search objects, and then in determining the amount of amplitude intervals N consider only the most significant amplitude intervals.

2. The method according to claim 1, characterized in that the area of search objects is determined by two one-dimensional even filtering by rows and columns, consistent with the size of the object in rows and columns, respectively, and two-dimensional odd filter in rows and columns, consistent with the size of the object in rows and columns, respectively, the formation of the criterion function in accordance with the results of the even and odd filter and the centers of the areas of search objectpooling background of such centres of the areas of search, in which the merit function is above a threshold level.

3. The method according to claim 2, characterized in that the two-dimensional odd and even filter the walkie-talkie in rows and columns is performed recursively.

4. The method according to claim 1, characterized in that to account for the most significant amplitude intervals when estimating N in square box scan first form of the brightness histogram by counting the frequency of contact of the brightness values of the image in the appropriate amplitude intervals, and then consider only the amplitude of the intervals where the values of the histogram exceeds a threshold level.