ADAPTIVE AUTONOMOUS NAVIGATION AND TARGET TRACKING SYSTEM FOR DEEP-LEARNING BASED UAVS

Abstract

Relevance. This paper presents an autonomous control system for unmanned aerial vehicles (UAVs) based on Robot Operating System (ROS) for efficient real-time object detection and tracking under limited computational resources. The system integrates a modified YOLOv4 architecture for high-speed object detection and the SiamMask algorithm for continuous target tracking. Flight trajectory control is implemented through a PID controller ensuring stable automatic target following. The key novelty of the research lies in developing a methodology for optimizing neural network models for embedded systems with critical computational power constraints. The proposed approach enables adaptation of modern deep learning architectures for operation on single-board computers and embedded GPUs without significant loss of detection accuracy.

The system is based on a combination of computer vision methods and machine learning algorithms. Convolutional neural networks trained on extended datasets provide robust object recognition under various weather and lighting conditions. The tracking module employs optical flow and Kalman filter with predictive algorithms for trajectory forecasting. The system is resistant to noise, interference, and partial target occlusion.

Experimental validation was conducted in indoor and outdoor environments using real video data and simulations. Test scenarios included tracking of vehicles, people, and dynamic objects under various conditions.

The developed architecture expands the application capabilities of autonomous UAVs in monitoring tasks, search and rescue operations, and industrial inspection, where weight-size and energy characteristics of onboard equipment are critical.