METHOD FOR DETECTION AND TRACKING OF SMALL AERIAL TARGETS BASED ON STABILIZED INTER–FRAME DIFFERENCE WITH ADAPTIVE LOCAL THRESHOLD SEGMENTATION

Abstract

The article addresses the highly relevant scientific and practical problem of detecting and multi–track tracking of small aerial targets (Unmanned Aerial Vehicles, UAVs) in a video stream captured continuously from the onboard camera of another observing UAV. The broad integration of computer vision technologies onboard drones is severely hindered by strict Size, Weight, and Power (SWaP) constraints. This structural limitation makes the deployment of modern, resource–intensive deep neural networks (such as YOLO or Faster R–CNN) virtually impossible without specialized graphics processing units (GPUs). Additional complicating factors in real–world environments include the extremely small sizes of the targets (often less than 15 pixels), the inherently low contrast of objects against the sky or ground background, and the continuous movement of the camera itself (ego–motion), which inevitably creates a global background shift and motion artifacts.

The primary aim of this research is to develop an efficient, computationally lightweight detection pipeline for small moving targets, strictly suitable for real–time deployment on low–power onboard Edge systems running on standard central processing units (CPUs). To achieve this ambitious goal, a modified computer vision methodology based on inter–frame difference is proposed. The newly developed method seamlessly integrates Lucas–Kanade affine image stabilization utilizing RANSAC, adaptive local threshold segmentation based on local mean and standard deviation matrices, an Exponential Moving Average (EMA) background model to gracefully smooth out slow lighting changes, and a continuous temporal voting concept featuring smooth exponential decay to aggressively filter out random noise and artifacts. The strategic use of local spatial statistics allows the complete abandonment of fixed global thresholds, thereby maintaining a high sensitivity to extremely small targets while effectively suppressing false positives (FP) in high–contrast or noisy areas.

The comprehensive experimental validation of the developed method was conducted on a challenging dataset consisting of 15 video clips with a total duration of over 14,000 frames, thoroughly annotated in the standardized MOT 1.1 format. The extensive testing results demonstrated that the proposed method provides an excellent average processing speed of 38.5 frames per second, which comfortably satisfies real–time operational requirements (exceeding the baseline of 25 FPS). The overall average F1-score metric reached 0.304 across the dataset. Furthermore, a detailed per–object analysis conclusively revealed that detection quality heavily correlates with the target's pixel size and motion characteristics: detection completeness (Recall) critically plummets for objects smaller than 15 pixels and at relative speeds exceeding 10 pixels per frame. The study solidifies the high potential of the proposed algorithmic pipeline for autonomous onboard drone computers, as it operates efficiently and robustly without the need for hardware graphics acceleration, effectively reserving critical power and computational resources for core flight management tasks.

Keywords: unmanned aerial vehicle, small moving target detection, image stabilization, inter–frame difference, adaptive threshold segmentation, temporal voting, multi–track tracking.