🧭
The complementary multimodal inputs enable accurate pose estimation, while fusing SLAM-based global camera poses with IMU data ensures low Root Position Error and stable global trajectory tracking, reducing visual drift and mitigating errors inherent in IMU-only methods.
⭐ Abstract
We present EveryWear, a lightweight and practical on-body human motion capture system that uses only everyday consumer wearables—a smartphone, smartwatch, earbuds, and smart glasses with built-in forward and downward cameras—without requiring any explicit calibration. To support real-world training and benchmarking, we introduce Ego-Elec, a 9-hour MoCap-annotated dataset spanning 56 daily activities across 17 indoor and outdoor environments. Our multimodal teacher-student framework, trained entirely on real-world data, achieves more accurate pose estimation while avoiding the sim-to-real gap.
🖼️ Dataset
🏷️
Pie Chart: Distribution of Environments. Bar Chart: Distribution of Activity Types.
👓 System
⚙️
Our system comprises smart glasses, smartphone, smartwatch, and earbuds. The signals collected at different frequencies are then aligned and downsampled.
💡 Method
🔎
We introduce a teacher-student framework that takes as input multi-view egocentric images from smart glasses together with inertial measurements, to estimate full-body human motion, including global translation and joint rotations over time. In addition, we incorporate a standard SLAM module to recover global head position.
📊 Results
🧮
Performance of EveryWear and other Baselines on Ego-Elec.
📦
Compared to other baselines, our model achieves higher pose estimation accuracy and lower Root PE, demonstrating superior performance across both positional and pose-related metrics.
📝 BibTeX
@inproceedings{zhu2025human,
title={Human Motion Estimation with Everyday Wearables},
author={Zhu, Siqi and Li, Yixuan and Li, Junfu and Wu, Qi and Wang, Zan and Ma, Haozhe and Liang, Wei},
booktitle={},
year={2025}
}