МЕТОД КЛАСТЕРИЗАЦИИ ЛЕТЯЩИХ ОБЪЕКТОВ ПО ВИДЕОПОТОКУ НА ОСНОВЕ НЕЙРОННЫХ СЕТЕЙ

Abstract

The paper presents a method of using a neural network for clustering objects flying against the sky background. Such a method is proposed to be used for the tasks of the detection or assignment to a certain cluster, as well as for quickly determining the danger of an object flying through monitoring zone. To solve this problem, a neural network is used, consisting of a convolutional network (determining the static features of an object, for example, shape and color) combined with a recurrent network (determining the dynamic features of an object, for example, the movement of propeller blades, flapping wings), the results of which are fed to the Kohonen map, where the clustering of the object occurs. Examples of input and output data of the method are shown, as well as ways to interpret the output data (Kohonen map) to quickly determine the danger of a flying object. Methods for preliminary processing of input data are given. The article also outlines the basics for the architecture of recurrent and convolutional neural networks used in the presented method. A training algorithm and a set of data used in neural network training are described. Studies have been carried out and the results of the method are presented. An estimate of the effectiveness of the method for video analytics of flying objects taken from a video stream is given.