Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Spectral-Temporal Fusion Representation for Person-in-Bed Detection

Xuefeng Yang, Shiheng Zhang, Jian Guan, Feiyang Xiao, Wei Lu, Qiaoxi Zhu

TL;DR

This work tackles bed occupancy detection using mattress-embedded accelerometer signals by introducing a spectral-temporal fusion feature representation, where $\mathbf{H} = \mathcal{F}(\mathbf{H}_s, \mathbf{H}_t)$ combines Mel-filter spectral features and a temporal spectrum from $\text{TgramNet}$ via a CNN fusion module. Separate heads are designed for segmented detection (Track 1) and streaming detection (Track 2), employing a MobileFaceNet-based projector, mixup data augmentation, and an IoU loss to improve boundary localization. The approach achieves top performance on ICASSP 2025 benchmarks, obtaining 100.00% in Track 1 and 95.55% in Track 2, demonstrating robustness to variability and disturbances with practical implications for smart-home sensing. These results highlight the value of integrating spectral and temporal cues for accurate, real-time bed occupancy monitoring using inexpensive mattress-embedded sensors.

Abstract

This study is based on the ICASSP 2025 Signal Processing Grand Challenge's Accelerometer-Based Person-in-Bed Detection Challenge, which aims to determine bed occupancy using accelerometer signals. The task is divided into two tracks: "in bed" and "not in bed" segmented detection, and streaming detection, facing challenges such as individual differences, posture variations, and external disturbances. We propose a spectral-temporal fusion-based feature representation method with mixup data augmentation, and adopt Intersection over Union (IoU) loss to optimize detection accuracy. In the two tracks, our method achieved outstanding results of 100.00% and 95.55% in detection scores, securing first place and third place, respectively.

Spectral-Temporal Fusion Representation for Person-in-Bed Detection

TL;DR

This work tackles bed occupancy detection using mattress-embedded accelerometer signals by introducing a spectral-temporal fusion feature representation, where combines Mel-filter spectral features and a temporal spectrum from via a CNN fusion module. Separate heads are designed for segmented detection (Track 1) and streaming detection (Track 2), employing a MobileFaceNet-based projector, mixup data augmentation, and an IoU loss to improve boundary localization. The approach achieves top performance on ICASSP 2025 benchmarks, obtaining 100.00% in Track 1 and 95.55% in Track 2, demonstrating robustness to variability and disturbances with practical implications for smart-home sensing. These results highlight the value of integrating spectral and temporal cues for accurate, real-time bed occupancy monitoring using inexpensive mattress-embedded sensors.

Abstract

This study is based on the ICASSP 2025 Signal Processing Grand Challenge's Accelerometer-Based Person-in-Bed Detection Challenge, which aims to determine bed occupancy using accelerometer signals. The task is divided into two tracks: "in bed" and "not in bed" segmented detection, and streaming detection, facing challenges such as individual differences, posture variations, and external disturbances. We propose a spectral-temporal fusion-based feature representation method with mixup data augmentation, and adopt Intersection over Union (IoU) loss to optimize detection accuracy. In the two tracks, our method achieved outstanding results of 100.00% and 95.55% in detection scores, securing first place and third place, respectively.
Paper Structure (9 sections, 2 equations, 1 figure, 1 table)

This paper contains 9 sections, 2 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Spectral-Temporal Fusion-Based Feature Representation
  4. Segmented Detection Head for Track 1
  5. Streaming Detection Head for Track 2
  6. Results
  7. Dataset and Metric
  8. Results and Analysis
  9. Conclusion

Figures (1)

  • Figure 1: Structure of our models for person-in-bed detection. (a) Spectral-temporal fusion-based feature representation. (b) Segmented detection model and (c) streaming detection model, both based on (a) but operating on different tracks.