Research map / 3D vision / Human pose estimation
UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UMotion, an uncertainty-driven UKF-based online state estimation framework fusing six body-worn IMU-UWB sensors with a learned shape estimator and an LSTM pose estimator for 3D human shape and pose estimation.
Problem
Sparse-IMU 3D human motion estimation suffers from pose ambiguity, IMU drift, UWB occlusion, and poor adaptation to diverse body shapes.
Approach
Treats inter-node UWB distances as core system state, estimates body shape from height, weight and inter-distances, and tightly couples IMU, UWB and pose-estimator outputs (with uncertainty propagated via an unscented transform) in a UKF so that pose uncertainty refines sensor inputs and vice versa.
Key results
- On TotalCapture, UMotion improves over the IMU-only SOTA PNP by 32.4% in angular error and 9.4% in mesh error.
- Versus distance-augmented UIP, UMotion reduces positional error by 21.5% on TotalCapture and 35.0% on DIP-IMU.
- State estimator reduces mean inter-sensor distance error from 9.20 cm to 2.42 cm on TotalCapture with synthesized LOS noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add UWB distance constraints and the state-estimation framework to IMU pose estimation | UMotion · TotalCapture | Angular Error: 10.45 -> 7.06 | Helped |
| Add UWB distance constraints and the state-estimation framework to IMU pose estimation | UMotion · DIP-IMU | Positional Error: 4.97 -> 3.38 | Helped |
| Use UMotion instead of UIP | distance-augmented pose estimation · TotalCapture | Positional Error: 5.49 -> 4.46 | Helped |
| Use UMotion instead of UIP | distance-augmented pose estimation · DIP-IMU | Positional Error: 5.05 -> 3.38 | Helped |
| Use UMotion instead of UIP | distance-augmented pose estimation · UIP | SIP Error: 24.12 -> 25.69 | Hurt |
| Use UMotion instead of UIP | distance-augmented pose estimation · UIP | Positional Error: 10.65 -> 10.33 | Helped |
| Use the predicted body shape rather than mean shape (ablation) | UMotion shape reconstruction · TotalCapture | Positional Error: 5.15 -> 4.34 | Helped |
| Add an intermediate joint-position estimate from sensor measurements before pose regression (ablation) | UMotion pose estimator · TotalCapture | SIP Error: 10.76 -> 12.15 | Hurt |
| Add an intermediate sensor global-position estimate from sensor measurements before pose regression (ablation) | UMotion pose estimator · TotalCapture | Positional Error: 4.46 -> 4.96 | Hurt |
| Add an intermediate sensor relative-position estimate from sensor measurements before pose regression (ablation) | UMotion pose estimator · TotalCapture | Positional Error: 4.46 -> 5.44 | Hurt |
| Fuse IMU, UWB, and pose data instead of using raw distances (ablation) | UMotion state estimator · TotalCapture with synthesized LOS-related noisy distances | Distance Error: 9.20 cm -> 2.42 cm | Helped |
| Fuse IMU with UWB and pose data rather than leaving measurements unfiltered (ablation) | UMotion state estimator · TotalCapture with synthesized LOS-related noisy distances | Joint positional error | Helped |
Surprising findings
- Introducing an intermediate joint/position estimation layer increased errors, suggesting a simpler direct IMU-plus-distance architecture performs better.
Limitations
- Shape estimator needs adaptation to specific sensor configurations and conditions and may struggle with unusual body variations given limited training data.
- State estimator requires careful parameter tuning to reflect sensor characteristics.
- Pose estimator simplifies constraints, which may limit performance; no physics-aware module.
Experimental setup
Datasets: AMASS TotalCapture DIP-IMU UIP
Benchmarks: TotalCapture DIP-IMU UIP
Metrics: SIP Error Angular Error Positional Error Mesh Error
Baselines: DIP TransPose TIP PIP PNP TIP-D PIP-D UIP
Builds on: SMPL PIP UIP DIP AutoGluon
Closest papers by meaning
- Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors (CVPR 2024 · similarity 0.87)
- DiffusionPoser: Real-time Human Motion Reconstruction From Arbitrary Sparse Sensors Using Autoregressive Diffusion (CVPR 2024 · similarity 0.86)
- Spatial-Related Sensors Matters: 3D Human Motion Reconstruction Assisted with Textual Semantics (AAAI 2024 · similarity 0.85)
- Loose Inertial Poser: Motion Capture with IMU-attached Loose-Wear Jacket (CVPR 2024 · similarity 0.85)
- HiPoser: 3D Human Pose Estimation with Hierarchical Shared Learning at Parts-Level Using Inertial Measurement Units (AAAI 2025 · similarity 0.85)
- EnvPoser: Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.