Large Model-Driven Real-Time Object Detection in High-Definition Video Surveillance Over Wireless Networks

Abstract

This paper proposes LMDOF-SIOGAN, a novel large model-driven object detection framework that integrates a Swarm Intelligence-Optimized Generative Adversarial Network (SIOGAN) with an adaptive large model detection pipeline. The proposed framework introduces two key modules: (1) a VE-GAN-based visual enhancement module (VE-GAN), which leverages adversarial learning and perceptual supervision to restore semantic integrity from degraded video frames; and (2) a Swarm Intelligence-Aided Scheduler (SIAS), which dynamically optimizes the detection pipeline based on real-time network conditions and video quality assessments. Extensive experiments were conducted on public datasets VisDrone2021 and UAVDT, under simulated wireless video surveillance environments with multi-resolution streams and comprehensive network impairments. The results demonstrate that LMDOF-SIOGAN consistently outperforms state-of-the-art baselines including YOLOv8 and ApproxDet, achieving up to 80.3% [email protected] at 4 K resolution with 125 ms end-to-end latency, while maintaining superior robustness index (RI) and moderate computational overhead. Ablation studies further validate the critical contributions of the VE-GAN and SIAS modules to detection accuracy, robustness, and latency.