GENERALISED PROBLEMS OF COMPUTER VISION IN THE DECISION-MAKING TASKS BY AUTONOMOUS UNMANNED AERIAL VEHICLE SWARMS

Abstract

Computer Vision (CV) is one of the key technologies in the modern unmanned aerial vehicles (UAVs) industry because it allows us to get insights from the camera, recognizing different objects and so on. Although this is not a new technology, it is rapidly growing bringing up new abilities and features from implementation. Sometimes it is hard to implement new features due to different academic or industrial problems related to CV. Thus, the paper presents the results of comprehensive research on CV problems that should be considered when modelling and designing decision-making systems for both autonomous UAVs and their swarms. Attention was focused on aspects such as computational resource limitations, unstable environments, inter-UAV synchronisation, visual data exchange, scale variability, object orientation, adaptability to real-time task changes, limited field of view, etc. Among the analysed existing solutions, conclusions were drawn about the advantages of systems designed to solve the tasks of piloting UAVs and making decisions by autonomous UAV swarms, as well as their disadvantages, which must be taken into account when building new solutions. According to the research results, several problems of CG have been formalised that do not have a complete solution, namely, loss of accuracy of CG models when the environment changes, merging of local UAV maps, operation in conditions of "blind zones", monitoring of surrounding objects, coordination of CG models in the process of self-learning by a swarm of UAVs, obtaining knowledge using KM and assessing the uncertainty of models in new conditions when the model does not know what it does not know. Based on the analysis of the problems considered, requirements for building system solutions for autonomous UAV swarms were formed and directions for further research were proposed, in particular, in the areas of distributed computing, training UAV swarms, multi-agent systems with stereo vision, and retraining in real conditions.