Stochastic Latency Decomposition and Constrained Runtime Feasibility Analysis for Edge-Based UAV Surveillance Under Network-Denied Environments

In security and tactical surveillance applications, unmanned aerial vehicle (UAV) detection systems must provide both reliable recognition and stable real-time operation under communication-constrained conditions. However, remote server-based surveillance can suffer from unstable response times when the display or output path depends on a degraded network. This study formulates edge-based UAV surveillance under a network-denied operating condition as a stochastic latency-decomposition and constrained runtime-feasibility problem. The total system latency is decomposed into inference, processing, and display/I/O components, and an SSH X11-based lossless display-path proxy is used to examine how network-coupled output transmission can dominate the runtime path. In contrast, a Jetson AGX Orin-based edge implementation performs UAV detection, tracking, threat assessment, visualization, and output locally. A YOLO26-based reference detector accelerated with TensorRT and FP16 is evaluated using a high-resolution UAV dataset consisting of approximately 25,000 images from nine UAV classes. Five-fold cross-validation produced an mAP@0.5 of 0.7890 ± 0.0653. Runtime evaluation showed that the optimized edge system achieved 31.49 ± 2.49 FPS at SD resolution, satisfying the strict 30 FPS real-time condition, while HD resolution achieved 26.72 ± 1.31 FPS as a near-real-time high-detail mode. Under the SSH X11 proxy condition, the FHD runtime dropped to 4.85 ± 2.53 FPS with substantially increased display latency. These results indicate that real-time UAV surveillance depends not only on detector inference speed but also on execution architecture and display-path dependency, supporting the practical importance of network-independent edge deployment under communication-degraded conditions.