CC BY
5УДК 004.93'1
DEDICATED VIDEO ANALYTICS AS A TECHNOLOGY OF COMPUTER
VISION
Beknazarova Saida Safibullaevna DSc, Professor, Tashkent University of Information Technologies named after Muhammad Al-Khwarizmi,
E-mail: [email protected], Tel:+998903276666
Abdullaeva Ozoda Safibullaevna PhD, Professor of Namangan Engineering-Construction Institute, E-mail: [email protected], Tel: +998913549363
Аннотация: Маколада видеотах,лил усулларига багишланган - реал вакт режимида видеокамералардан ёки архив ёзувларидан олинган тасвирлар кетма-кетлигини тах,лил килиш асосида турли хил маълумотларни автоматлаштирилган кайта ишлаш учун компьютер куриш усулларидан фойдаланадиган технология куриб чикилган. Динамик объектларни аниклаш вазифаси деганда, кадрлар кетма-кетлигида тасвирнинг узгарувчан жойларини аниклаш ва ажратиш вазифаси очиб берилган. Шунга кура, маълум бир объектни топиш деганда маълум бир кидирув объекти билан баъзи ухшаш хусусиятларга эга булган бир ёки бир нечта аникланган динамик объектларни танлаш тушунилади. Объектлар алгоритмга мувофик танланади. Объектни кидириш жараёни Афин, пройектив бузилишлар, объектнинг бошка объектлар билан бир-бирига ёпишиши ва кабул килувчи (датчик) томонидан аникланади.
Аннотация: Статья посвящена видеоаналитике - технологии, использующей методы компьютерного зрения для автоматизированного сбора различных данных на основе анализа последовательности изображений, поступающих с видеокамер в режиме реального времени или из архивных записей. Под задачей обнаружения динамических объектов понимается задача обнаружения и выделения изменяющихся областей изображения в последовательности кадров. Соответственно, обнаружение определенного объекта означает выбор одного или нескольких обнаруженных динамических объектов, которые имеют некоторые сходные признаки с данным объектом поиска. Объекты выбираются в соответствии с алгоритмом. Процесс поиска объекта осложняется аффинными, проективными искажениями, перекрытием объекта другими объектами и помехами приемника (датчика).
Abstract: The article dedicated video analytics is a technology that uses computer vision methods for automated acquisition of various data based on analysis sequence of images coming from video cameras in the mode real time or from archival records. Under the task of discovery dynamic objects is understood as the task of detection and selection changing areas of the image in a sequence of frames. Accordingly, the detection of a certain object means the choice one or more detected dynamic objects that have some similar features with a given search object.
Features are selected according to the algorithm. Search process object is complicated by affine, projective distortions, overlapping object by other objects and receiver (sensor) noise.
Калит сузлар - усул, алгоритм, объектни кидириш, видео оким.
Ключевые слова — метод, алгоритм, поиск объекта, видеопоток.
Keywords— method, algorithm, searching an object, video stream.
Introduction
According to the system requirement, the algorithm should be based on the search for the key points of the object. The conducted studies have shown that the ASIFT (Affine Scale-Invariant Feature Transform) method is the most resistant to the criteria considered. ASIFT method is based on the SIFT (Scale-Invariant Feature Transform) method, which has a fast-acting analogue - the SURF (Speeded up Robust Features) method [1].
The SIFT method is the most resistant to the criteria considered, but has a high computational complexity. The SIFT algorithm is a complication of the SIFT method, which makes it possible to achieve stability to all affine transformations by modeling changes in camera tilt. According to the constructed functional model, the camera tilt changes are generated by the function, thereby the ASIFT method becomes redundant in finding singular points.
In SIFT, the key point is considered to be the local extremum in the scalable space of the Gaussian difference. In the fast-acting analog, the SURF method [2], the key point is the local extremum of the determinant of the Hesse matrix. In practice, the SURF algorithm allocates fewer key points on the object image, but has a high frame processing speed compared to the SIFT method [3]. To quickly find key points and calculate descriptors, it is proposed to use the SURF method as a basis.
Considering the constructed functional model, the stability to scaling in the system is achieved by performing the function on the image of the desired object; therefore, the SURF method has been upgraded: the search for key points is performed only on one octave. In this regard, the computational complexity of the upgraded method is reduced by я times, where s is the number of octaves.
Methods
Algorithms and methods for finding the intersection of descriptors
To find the intersection of two descriptor sets, the following approaches are actively used
today:
- RANSAC (RANdom SAmple Consensus) method;
- Kuhn - Mankres algorithm.
RANSAC - This is a general method that is used to estimate model parameters based on random samples. When compared, the model is a transformation matrix (homography). There are two sets of descriptors at the input of the algorithm. The scheme of work of RANSAC consists of repeated repetition of three stages:
1. Selection of points and construction of model parameters. From the input sets of descriptors, sets of fixed size are randomly selected without repetition. Based on the obtained sets, a transformation matrix is constructed.
2. Checking the constructed model. For each descriptor of the object image, a projection is located on the current frame and a search is performed for the closest descriptor from the set of descriptors of the current frame. The descriptor is marked as an outlier if the distance between the projection and the corresponding descriptor of the current image is greater than a certain threshold[3].
3. Replacement of the model. After checking all the points, it is checked whether the constructed model is the best among the set of previous models. As a result of using RANSAC, the best homography matrix is constructed. Having calculated the perspective projection of a set of object image descriptors, it is enough to pass through all the correspondences obtained during the iteration and check whether the corresponding descriptor of the current frame is close enough to the projection of the object image descriptor. If it is not, then the pair is discarded [4].
According to [5] for one model, the computational complexity will be O(n) however, in practice, the results are unacceptable for use due to the large number of possible errors. There are modifications of the RANSAC method. For example, the G-Linkage algorithm and the kernel adaptation algorithm, which allow finding pairs with fewer errors, but with computational complexity O(n)2 [6]. The Fischer scaled Compressed Vector algorithm with RANSAC (SCFV (Single-chain variable fragment)-RANSAC) [7] similarly has fewer errors due to additional processing of the set of descriptors for matching.
The Kuhn-Mankres algorithm
The task of matching descriptors can be represented as an assignment task. We interpret it into a graph form. Let the mask parameters (descriptors) be the vertices of the graph, and the values of the vertex similarity measure are the edges of this graph. The complexity of the original algorithm is . O(n)4 To solve the problem by the Kuhn-Mankres method, it is necessary to add new virtual vertices of the graph, which will be infinitely removed from other vertices. ThenK(n,n) [W] - weighted graph with fractions X and Y. The output of the method is the set of
edges of the optimal match P in this column.
The Kuhn - Mankres method can be represented as the following sequential operations:
1. Set to K(n,n)[W] arbitrary acceptable markup f and find a subgraph of equalities G(W, f).
2. Using the Hungarian algorithm to find the maximum match P in the graph and a lot of F free relatively P share vertices X.
3. If F = 0, finish the job.
4. Find all alternating chains in the graph, starting in F, put S and T equal to the set of all vertices of the fraction X(accordingly, the shares Y), met in these chains.
5. If in T there are no free vertices, put
A = min(x; e S,yt e y/T){f(xi ) + f(y, )-wiy } (1)
Where f(x) = f(x) - A for everyone x e S, f(y) = f(y) + A for everyone y eT, find a new graph G(W,f) and go back a step 4.
6. Increase P, by repainting the magnifying chain found, and go back a step 3 [8].
Algorithm for limiting the search area of an object in the frame
The algorithm for limiting the object search area evaluates the scale of the object image by the descriptors of key points according to the following scheme:
1. Find for each key point of the frame the closest match from the set of projectively distorted images of the sample.
2. Remove from further consideration the key points of the frame that have the value of the proximity measure below the threshold.
3. For each remaining key point of the frame, build a rectangular area. The coordinates of the selected area in the image are determined by the coordinates of the corresponding key point in the distorted image of the sample.
4. From the set of key points for further analysis, leave only those whose rectangular areas have intersection areas with other rectangular areas less than half the area of the rectangular area of the considered key point [9].
Results
The algorithm is presented in more detail in the form of a flowchart in Figures 1 and 2, constructed on the threshold of the proximity measure of descriptors.
The measure of proximity between the frame descriptors and the sample image is calculated by the Bhattacharya coefficient [9]:
where a,b - dimension vectors n,p- measure of proximity, p e [0,1]
The proposed algorithm has less computational complexity than the algorithms of RANSAC and Kuhn-Mankres. However, the disadvantage of this approach is to determine the Threshold value.
(2)
Figure 1-block-circuit algorithm limited areas requested in Cadre
The identification algorithm should determine whether the area on the frame is an image or part of the image of the object. To do this, the algorithm must find the parameters of the window on the frame by the found areas obtained based on the comparison of local image features - key points. Let the identification algorithm find the object by an elliptical window. The algorithm suggests using a method based on the global property of the image. One of the most common global characteristics is the color histogram. The color histogram is calculated quickly, however, the spatial arrangement of pixels is not taken into account when calculating. It is proposed to enter the point color values with a certain weight: the closer the point is to the center of the window, the greater its weight. This is also necessary so that small window offsets lead to small changes in the mapping error:
K(x) = {1 - x2, |x| < 1,0X ^1 (3)
Thus, the color of the pixel x will be entered into the color histogram with a certain weight K(x).
It is proposed to use gradient descent to localize the object. It is proposed to use the Bhattacharya coefficient as a criterion of similarity.
Yes Do/and к intersect
rectanEular areas ^^
Calculate the areas of and Prenons respectively, as well as the area of intersection of 5 j and к fseions
wt := wc - 1
_ f
Figure 2- Flow diagram of the process of removing rectangular object search areas
Image area identification algorithm
Gradient descent is used to solve the problem of finding a local minimum. For a stable identification process to minor color changes and to reduce the size of the histogram, quantization of the color histogram values is performed.
Discussion
The four-parameter search for the object image window by gradient descent method consists of the following steps:
1. Ask pth (minimum similarity threshold value), imax (maximum number of
iterations), a0 u h0 (the lengths of the semi-axes of the ellipse).
2. Calculate the normalized frequency vector of intensity values (histogram) hisf^, in the elliptical region x0, y0, h0 ,®0
3. Calculate the measure of similarity using the Bhattacharya coefficient:
Po = XV hiStrhfbhiSt0b (4)
beB
Where hist rhfb - histogram of the object image to search for, i> - histogram step, b e B.
4. i = 1
5. while p.=1 < pth and i < /max perform:
5.1 Calculate the gradient gradp_(i-1):
gradvpt_i = (%!,%1, %) (5)
ox oy ow oh
5.2 Calculate k^:
Ki K 1 - P.-i(xi-i,y-i,hi-1,^-1) (6)
IgradPi.J2
5.3 Calculate the step value:
Axt, Ay, Ah., A® = k.-jgradpi-j (7)
5.4 Change Ellipse parameters:
xi, yi, hi, ®i = Axi, Ay., Ahi, + xi-i, y.-i, hi-i, ®i-1 (8)
5.5 Calculate the relative histogram hist(i) elliptical area xt,y,ht
5.6 Calculate the similarity measure pi:
pt =X^hist rhfbhistib (9)
beB
5.7 i = i +1, go to step 5. 6. Stop. Conclusion
Initial values w0 and /i0 are taken from the parameters of the considered area P, obtained after the algorithm for limiting the search area of the object in the frame.
The color histogram is calculated by the color-difference components U h V color space
YUV
U = -0.14713R- 0.28886G+ 0.436B+128 V = 0.615R- 0.51499G-1.0001B+128 where R, G, B are 8-bit color values.
Such a histogram has a smaller size, unlike a histogram consisting of three components of the RGB space, and such a histogram is more resistant to changes in the brightness component in the image.
Gradient descent, applied to determine the parameters of the four parametric model, has computational complexity 0(n2). As soon as the value of the color histogram matching criterion in the identification method becomes higher than a certain threshold pth, or the number of iterations exceeded the maximum value imax, the identification process is terminated. The image of an object is considered found on the frame if the minimum value of the similarity threshold is exceeded pth
REFERENCE
1. Guoshen, Yu. ASIFT: An Algorithm for Fully Affine Invariant Comparison, Image Processing On Line /Yu. Guoshen, M. Jean-Michel//Image Processing On Line. - 2011 - №1. http://www.ipol.im/pub/art/2011/my-asift/article.pdf.
2. Kenneth, D.-H. A Practical Introduction to Computer Vision with Open CV/D.-H. Kenneth. - Ireland: Trinity College Dublin, Ireland, 2014, - 234 p. - ISBN 978-1-118-84845-6.
3. Bernatovich A. S. An active experiment in the identification of functional systems for the operational implementation of simulation-type models / / Cybernetics.- 1983. - No.1.- pp. 99104.
4. Ayvazyan S. A., Bukhstaber V. M., Enyukov I. S., Meshalkin L. D. Applied statistics. Classification and dimension reduction. Moscow: Finance and Statistics, 1989. 608 p.
5. Anderson D. Tannehill G. Pletcher R. 1990 Vychislitelnya gidromehanika i petloobmen (Moscow: Mir) Moscow: Nauka)
6. Sedova N., Sedov V., Bazhenov R., Karavka A., Beknazarova S. Automated Stationary Obstacle Avoidance When Navigating a Marine Craft //2019 International MultiConference on Engineering, Computer and Information Sciences, SIBIRCON 2019; Novosibirsk; Russian Federation; 21 October 2019
7. Beknazarova S., Mukhamadiyev A.Sh. Jaumitbayeva M.K.Processing color images, brightness and color conversion//International Conference on Information Science and Communications Technologies ICISCT 2019 Applications, Trends and Opportunities. Tashkent 2019N.
8. Beknazarova S., Mukhamadiyev A.Sh. Park Insu, Adbullayev S. The Mask Of Objects In Intellectual Irrigation Systems//International Conference on Information Science and Communications Technologies ICISCT 2020 Applications, Trends and Opportunities. Tashkent 2020.
9. Beknazarova S., Sadullaeva Sh., Abdurakhmanov K, Beknazarov K. Nonlinear cross-systems of numerical simulation of diffusion processes//International Conference on Information Science and Communications Technologies ICISCT 2020 Applications, Trends and Opportunities. Tashkent 2020.
