INTELLIGENT COMPUTER VISION SYSTEM FOR UNMANNED AERIAL VEHICLES FOR MONITORING TECHNOLOGICAL OBJECTS OF OIL AND GAS INDUSTRY

The relevance of the research is caused by the necessity to develop modern computer vision systems for monitoring hazardous technological objects of oil and gas industry. The main aim of the research is to develop the intelligent computer vision system for unmanned aerial vehicles, which allows monitoring dangerous technological objects and analyzing the monitoring data in real-time on the board of the unmanned aerial vehicle. Objects: the concept of construction of intelligent computer vision system; new architectures of convolutional neural networks hardware-based using field programmable gate array; the method of unification of computing blocks and ways of parallel calculation in hardware-based convolutional neural networks; algorithms of error-correction encoding and decoding data for exchanging message between ground and airborne components of the intelligent computer vision system. Methods: methods of detection and classification objects in images using convolutional neural networks; convolutional neural network deep learning methods; methods of designing software and hardware systems. Results. We have been analyzed the current state of research in the field of monitoring hazardous technological objects of the oil and gas industry and developed the concept of construction of intelligent computer vision system for unmanned aerial vehicles for monitoring dangerous objects. The idea of analyzing the images, obtained at monitoring of technological objects and surrounding areas, directly onboard of the unmanned aerial vehicle in real time was the base in this concept. Moreover, it is shown that the use of hardware-based convolutional neural networks for providing such analysis in real time is required. The authors developed the convolutional neural networks architectures for computer vision system from promising subclasses LeNet5 and YOLO and proposed the algorithms of error-correction data encoding/decoding for messages exchanging between these components, considering the specifics of ground and air-borne components. The authors developed the original method of organizing calculation in hardware-based convolutional neural networks using field programmable gate array, which differs from the known ones by using the unified computing blocks and new ways of parallel calculation in layers in these convolutional neural networks. They proposed the architecture of computing device of the unmanned aerial vehicle which includes the blocks of the hardware-based convolutional neural networks and the data encoder/decoder. This device is based on the Altera Cyclone V SX system-on-a-chip. The paper demonstrates the first results of studying the device efficiency. The authors developed the software for the ground component of the computer vision system.