Research into technologies and development of artificial intelligence-based device control system

Abstract

Analysis of recent studies and publications. Among the existing solutions, commercial ones can be distinguished, which are more focused on the tasks of sports racing, extreme aerial photography and cinematography and in most cases use satellite navigation. In today's realities, with the active development of this direction, there are several special platforms with proposed solutions and software from the manufacturer (for example, the VGI-9 [1] and Zir System [2] platforms).

Purpose of research. The purpose of this research is to analyze machine vision and object detection methods, develop and implement an intelligent FPV device control system based on artificial intelligence, which is capable of performing automatic detection, tracking and tracing of selected objects in real time. It is planned to deploy the system at the edge level of the device with subsequent direct installation of the appropriate software.

Presentation of the main research material. The main requirements for the control system are performance, accuracy and resource efficiency. The system must provide video stream processing with a frequency of at least 30 displayed frames per second (Frames Per Second, FPS) on the target hardware platform Orange Pi 5.

System architecture. The developed device control system is based on the principles of modular architecture with a clear separation of responsibilities between components. The system operates on the principle of pipeline processing using multi-threading to minimize latency.

Practical testing of the system. The YOLOv11 model family is represented by five variants of different sizes from nano to extralarge.

A combined approach was used to train the object detection model: the model is trained on the large open VisDrone dataset [9], later the model can be further trained on user data that is specific to a particular task.

The test bench consists of an Orange Pi 5 Max device and a 13MP camera. In real conditions, the system requires settings in Headless mode with video output to the framebuffer (a memory area for temporary storage of the image before displaying it on the screen).

Experimental tests of the developed device control system were conducted to verify the operability of key components and assess their performance. The tests were performed on a static stand without physical integration with the device.

Conclusions. In the course of the research, an intelligent FPV device control system based on artificial intelligence was developed and implemented, capable of performing automatic detection and tracking of objects in real time on the Orange Pi 5 edge device with NPU RK3588. A modular multi-threaded architecture was designed with the integration of the MAVLink communication protocol with the ArduPilot flight controller and PID controllers for the autonomous target pursuit mode were implemented. Experimental testing confirmed the system's operability on a static stand, but revealed performance limitations.