Behavior-driven temporal risk modeling for real-time drowning detection using drone imagery
Bao T. Nguyen · Array · 2026
Drowning remains a major public safety challenge in coastal environments, where large surveillance areas, dynamic water conditions, and limited lifeguard resources hinder timely detection and rescue. Unmanned aerial vehicles (UAVs) provide a scalable platform for maritime monitoring. However, existing vision-based drowning detection approaches typically formulate the task as frame-level appearance classification or rely on computationally intensive spatiotemporal action recognition models. These formulations often struggle with visual ambiguity, transient environmental disturbances, and strict real-time constraints, limiting their practical deployment in safety-critical settings. In this paper, we propose RIDE (Risk-aware Intelligent Drowning dEtection), a behavior-driven framework that reformulates drowning detection as a temporal risk inference problem based on abnormal human motion dynamics observed from aerial video. Rather than treating drowning as a static visual category, RIDE decouples human localization from drowning inference and models drowning as a continuous risk evolution process, enabling early and stable warning under visually noisy coastal conditions. The proposed framework integrates lightweight human detection, short-term temporal motion encoding using a lightweight two-layer 1D convolutional temporal encoder ( Conv1D temporal encoder ), and risk-aware aggregation over bounded temporal windows. This design explicitly avoids heavyweight 3D convolutional architectures, allowing RIDE to operate as a low-order causal temporal estimator optimized for early detection under noisy observations and strict latency constraints. By aggregating motion-derived risk signals over short horizons, the framework effectively balances temporal smoothing and responsiveness, suppressing false alarms caused by transient disturbances such as waves, reflections, and swimmer splashes. Experiments on drone-based maritime datasets demonstrate that RIDE achieves F1-scores of up to 84.5% and AUC of 0.88 , outperforming detection-based and frame-level baselines. Moreover, RIDE reduces the average Time-to-Detection by 3.2 s, providing earlier alerts while maintaining real-time performance at 38 FPS with low false-alarm rates. Overall, these results demonstrate the effectiveness of behavior-driven temporal risk modeling for UAV-assisted drowning early warning and provide a strong foundation for future validation in operational coastal surveillance systems.