Neural network detection of group test objects by hyperspectral markers during external trajectory measurements

Abstract

The task of detecting elements of group fl ight test objects has been actualized. Its relevance is based on the growing need to provide external trajectory measurements when testing samples of advanced multi-agent systems, the elements of which are aircraft. An analysis was made of the existing algorithms for detecting objects used in modern optoelectronic stations. Based on the results of the analysis, it was concluded that these algorithms do not meet the requirements for testing group objects. As a problematic issue in their application is the complexity of the selection of objects against the atmospheric background. To ensure the detection of objects, it is proposed to use additional highly informative features that refl ect the physical nature of the observed objects. Such features are the features of the shape of the spectrum of light refl ected from various materials. In order to identify them, it is proposed to conduct a preliminary hyperspectral analysis of test objects with subsequent reduction in the data dimension. To eliminate errors that occur during rotation and partial overlap of objects, it was proposed to use neural network methods, in particular, the YOLO v2 neural network detector. A variant of data organization for its training is proposed, as well as a detector architecture that provides high accuracy and rate of information processing. These characteristics of the detector were confi rmed experimentally on the basis of images exported from the model of the session of external trajectory measurements. A procedure for implementing neural network detection of group test objects was proposed. The results obtained are relevant for solving the issues of hardware and software for tracking optoelectronic systems.