IoT UWB-radar array system with two-stage multi-task deep network for bed occupancy and posture surveillance via spatial echo feature map and cross-modal visual explainability

Abstract

Accurate, unobtrusive in-bed occupancy-posture IoT monitoring is critical both for timely detection of “missing patient” events”, reducing false alarms in nocturnal surveillance, and informing sleep posture-related health risk management. This paper presents a novel two-stage, multi-task, multiple ultra-wideband (UWB) radar IoT system for joint in-bed occupancy detection and fine-grained sleep posture classification, using an array of eight UWB radars. The technical novelty centers on three mechanisms: a View-Weighted Multi-Radar Network (VWM-Net) that adaptively fuses single-radar echo maps into a Spatial Echo Feature Map (SEFM); an integration of DenseNet and ConvNeXt2 (DCNX2); and a cross-modal generative model that translates SEFMs into human-interpretable depth posture images. Specifically, a UCYC-GAN is proposed by integrating a U-Net structure with a CycleGAN generator, and its visual utility is compared against NICE-GAN and BDBM. The system is trained and validated on a cohort of 100 participants across 10 fine-grained postures and 4 blanket conditions, supplemented with a life-sized dummy and cluttered objects to model non-human occupancy and environmental distractors. In the first stage, a dedicated classifier distinguishes humans from non-human entities with 99.09% accuracy. In the second stage, the multi-head DCNX2 backbone performed multi‑task classification, achieving 93.75% accuracy for 10 fine‑grained postures, 98% for four coarse‑grained postures, 88.25% for blanket‑coverage types, 98.13% for blanket presence, and 95.13% for participant gender. The overall can effectively mitigate false positives from environmental clutter, advances non-contact in-bed occupancy-posture surveillance, and, through a cloud-enabled web dashboard and APIs, supports real-time visualization and retrospective data analysis.