Применение нейронных сетей для классификации объектов в видеоданных в задачах автоматического мониторинга состояния биологических объектов

Abstract

A steady trend towards the development of technologies in everyday life and industry leads to their constant complication, to the introduction of new developments and methods, new ways of acquiring and processing information. Video surveillance systems have become an integral part of everyday life. Cameras are installed in apartment buildings, public places, even in apartments and private houses. However, one of the most important areas where computer vision has found significant application is the observation of biological objects and analysis of their behavior and state, especially in the context of nature protection and automation of animal behavior research. With the growth of quality and quantity of video materials, the problem of accurate and efficient methods for classifying objects in real time becomes relevant, especially when it comes to automatic monitoring of animal condition. There can be no mistakes here, so it is necessary to use only the latest technology. This paper describes the main types of object recognition, considers popular modern neural network architectures, and performs a comparative analysis of some of them in relation to the problem to be solved. In the course of work using methods of additional processing and image annotators SuperVisely and VGG Image Annotator, a dataset with more than a thousand unique images for extraction of relevant characteristics of objects was formed, and experimental studies of the quality of object recognition on video using pre-trained known neural network models were carried out. The requirements to the initial data for effective solution of the problem of automatic monitoring and prediction of the condition of biological objects have been defined and formulated. It is shown that in order to avoid the appearance of blind zones in the animal habitat it is necessary to use a sufficiently large number of cameras placed with overlapping of the investigated space so that the monitoring objects were constantly in the field of visibility. This will subsequently make it possible to compile an overall high-resolution picture made up of images from all cameras and on the basis of the obtained picture to classify objects using artificial neural networks.