For the detection of living organisms under study, an algorithm for finding the average background is used, followed by morphological subtraction of the background from the original images processed by the Gaussian filter. After that, threshold binarization is performed. The resulting spots correspond to the positions of the studied organisms on each frame.

To correct errors associated with the superposition of the studied organisms in the frame projection and identify individual characteristics of animals, tracking is applied. For each organism with on the current frame, the vector of characteristics is calculated - (x,y,x1,y1,x2,y2), where (x,y) are the coordinates of the center of the organism, (x1, y1) are the projected velocities of the body on the plane of the frame in the corresponding axes, (x2, y2) - projections of accelerations on the plane of the frame in the corresponding axes. Further, on the next frame, similar vectors are calculated for all detected objects, after which a covariance matrix of Daphnia correspondence between frames is compiled. From which the movement of each specific object between frames is determined, as well as their merging/divergence.

As a separate case, objects are analyzed that cannot be represented as a key point, because have a sufficiently large frame coverage, and can have gradient transitions on the surface of their body. In such cases, a segmentation neural network is trained as a detector, which is capable of identifying areas of the frame where the organism under study is located.

To determine the health status of a group of animals, a neural network is used that is trained to determine the degree of aging of the population as a whole. The network architecture is a recurrent encoder-decoder, the purpose of which is to highlight the signs of aging expressed as a numerical vector.

The input of the neural network is the set of vectors [x, y, x1, y1, x2, y2] (which were previously mentioned) passed through an additional convolutional neural network of the backbone type, which serves to condense and generalise information about the position and movement of the object.

The neural network is trained by shooting objects throughout the entire life cycle, taking into account the removal of newborns, thus training takes place on data reflecting continuous aging.

We plan then to use dynamical differential equations from the project "Mathematical Physics for Aging Discovery" for vectors from received embedding to count aging biomarker.

At the moment, a study is being carried out on the fundamental possibility of using network-survey based on the following principles:

With known camera calibration parameters, the transformation from a point to a straight line is known and specified unambiguously. This means that the area of matching for each point can be limited only by the corresponding epipolar line in another image. Moreover, it is possible to establish a one-to-one correspondence between the epipolar lines in the two images.

For technical simplification of such a one-dimensional search, in practice, a special representation of the image is used, in which its epipolar lines coincide with the rows of pixels. Such an image is called rectified, and the process of obtaining it is rectification. Usually, the rectification of a pair of images consists in projecting them onto some plane parallel to the stereo base.

For tracking organisms in 3-dimensional space, we used an analog of epipolar lines - horizon lines, on rectified images from cameras mounted on the walls of the aquarium perpendicular to each other. Because in the case when the main axes of the cameras are perpendicular to each other, the epipolar lines of one camera generally degenerate into points in another camera. As for the horizon line, we will consider the reflection in each of the chambers of a plane parallel to the bottom of the aquarium, while the properties of rectification and the resulting convenience of tracking special points (in our case, moving organisms act as classical special points) will retain their applicability.