Uncertainty-aware background subtraction and semantic-guided denoising for adverse environments

Abstract

In edge computing scenarios such as unmanned aerial vehicle (UAV) patrol and vehicular perception, moving object extraction is highly susceptible to failure under low-light, inclement weather, and low-framerate conditions because of constrained sensor signal-to-noise ratio (SNR) and environmental interference. Existing deep learning models also struggle to satisfy the strict computational limits of edge payloads. To address these issues, this paper proposes a lightweight, purely CPU-driven framework that integrates background subtraction and video denoising. In the foreground perception stage, three-dimensional uncertainty metrics—capturing spatial texture, temporal motion, and illumination variation—are extracted online. An interpretable sensitivity matrix dynamically fuses multiple complementary detectors, suppressing false alarms triggered by low-light noise and dynamic backgrounds. In the video restoration stage, the binary mask generated from perception serves as a semantic prior: aggressive temporal-averaging denoising is applied to static background regions, while guided filtering is used for moving objects to preserve edge details, creating a mutual reinforcement loop in which detection guides denoising and cleaner frames facilitate robust perception. Evaluations on public benchmarks covering extreme low-light and adverse weather conditions show that the proposed method improves the detection F1 score by over 175% compared with traditional baselines in low-framerate scenarios and outperforms all evaluated classical algorithms in night scenes. It also achieves an overall denoising gain of 2.08 dB. Operating without GPU acceleration or offline training, the system maintains a single-frame latency of approximately 45 ms, making it suitable for real-time perception on resource-constrained edge platforms.