Real-Time Detection and Threat Assessment of Aerial Objects in High-Clutter Environments Using YOLOv4 and Kalman Filtering

Abstract

This study presents a real-time, modular system for the detection, classification, tracking, and evaluation of threats from airborne objects in visually cluttered aerial environments. The system incorporates the YOLOv4 object detection model along with Kalman filtering to ensure temporal consistency and accurate object tracking. A real-time risk evaluation module based on a lightweight, rule-based design carries out threat evaluation including object category, speed, heading, and proximity to defended areas. The algorithm was tested in MATLAB and tried various aerial video scenarios, including low-altitude drone video and simulated missile trajectories. Performance was assessed under both CPU-only and GPU-enabled runs, achieving up to 37 FPS on NVIDIA RTX-class hardware and approximately 22 FPS on Apple M4 CPUs with Metal acceleration. The results show outstanding classification accuracy (>90%) and robust real-time tracking across busy scenes. The findings confirm the appropriateness of the system for use in border monitoring, drone defense, and smart airspace surveillance.