Nocturnal Seizure Monitoring via Monocular Depth Estimation and 3D Point Cloud Analysis

Automated monitoring of nocturnal epileptic seizures remains a significant challenge, arising from the inherent conflict between the need for high-fidelity clinical data and the patient's right to privacy in domestic settings. Traditional RGB-based video monitoring is highly intrusive, legally sensitive, and frequently fails in low-light conditions. This paper proposes a novel, non-contact, privacy-preserving framework for home-based seizure detection using Monocular Depth Estimation (MDE). Specifically, we leverage the advanced Vision Transformer-based Depth-Anything-V2 (DAV2) architecture to transform standard 2D video feeds into structural 3D point clouds in real time. To eliminate perspective distortions from varying camera vantage points, we implement a robust Random Sample Consensus (RANSAC) plane-fitting algorithm for automated tilt correction, establishing a gravity-aligned baseline. Quantitative evaluations demonstrate that our DAV2-driven pipeline achieves a 34% reduction in Root Mean Square Error (RMSE) for structural mapping compared to existing state-of-theart monocular baselines, matching the localization precision of hardware-based Time-of-Flight (ToF) sensors while maintaining complete structural anonymization.