Development of an Accelerated Method for Calculating Streaming Video Data Obtained from UAVs

The problem of accelerating the processing of video streams captured from the unmanned aerial vehicle (UAV) is solved by the example of detecting smoke and fires (onboard) and finding target objects on panoramic images (on a ground computing platform). Due to the limited computing resources, all processing of video streams should be subdivided into tasks to be solved onboard and ground installations with operator control. It is proposed to use an algorithm for analyzing spectrographic textures based on the Euclidean-Mahalanobis metric. The use of the classifier of spectrographic textures provides finding it difficult to detect objects, which is unattainable for classifiers working with separate image points. It is proposed to use a ground computing facility with a neural network target recognition system with support for pipelining and parallelization. When implementing the processing of flows at ground stations, the advantage should be given to the methods of pipeline-parallel computations using cluster installations, which can be equipped with graphics accelerators; here an essential role is played not only by methods of accelerating computations but also by methods of automating the construction of computational schemes of problems due to a special graphical interface and modular organization of software. An experimental study in terms of speeding up computations in problems of recognizing objects and areas of interest in various computing systems has confirmed the correctness of the proposed approaches. The proposed methods and approaches together provide acceleration of computations by more than 10 times.