← Back to blogs

Sparse Inertial Motion Capture: IMU Models, Orientation, and Sensor Fusion

IMU-based pose estimation from sparse sensors

Human motion capture traditionally relies on marker-based optical systems (Vicon, OptiTrack) or dense sensor suits, both requiring controlled environments and expensive infrastructure. Sparse inertial motion capture is an alternative, using only 6 Inertial Measurement Units (IMUs) placed at key body locations to reconstruct full-body pose in unconstrained environments (Yi et al. 2021). Unlike vision-based methods, IMU-based approaches are immune to occlusion and work in any lighting condition, though they face challenges from sensor drift, noise, and the inherent ambiguity of mapping sparse measurements to full-body pose.

Early learning-based methods applied recurrent neural networks to model temporal dependencies in IMU sequences, but struggled with physical plausibility, producing motions with floating, sliding, or ground penetration artifacts. Physical Inertial Poser (PIP) (Yi et al. 2022) introduced physics-aware optimization to enforce ground contact constraints and biomechanical feasibility, combining neural kinematics estimation with physics-based refinement. Physical Non-inertial Poser (PNP) (Yi et al. 2024) extended this by modeling fictitious forces in non-inertial reference frames, addressing errors that arise when the pelvis (root) undergoes significant acceleration or rotation.

Calibration is a standing difficulty. Sensors must be precisely aligned to body segments, and magnetometer drift causes heading errors over time. Traditional calibration requires static poses (T-pose, N-pose), breaking workflow and failing when sensors shift. Transformer IMU Calibrator (TIC) (Zuo et al. 2025) achieves dynamic, implicit calibration by learning to estimate calibration matrices from diverse motion patterns, so that sensors can be put on and used without an explicit calibration step.

Multi-modal fusion enhances accuracy by combining IMU with complementary sensors. DiffCap (Pan et al. 2025) fuses sparse IMUs with monocular camera using diffusion models, where visual information provides dense constraints when available and the IMU stream keeps tracking during occlusion. BaroPoser (Zhang et al. 2025) incorporates barometric pressure from everyday devices (smartphones, smartwatches) to estimate height changes, enabling motion capture on non-flat terrain. Ultra Inertial Poser (Armani et al. 2024) adds ultra-wideband (UWB) ranging between sensors, providing absolute inter-sensor distances that reduce drift and jitter (97% jitter reduction reported on UIP-DB).

Recent work addresses practical deployment challenges. Loose Inertial Poser (Zuo et al. 2024) enables motion capture from sensors embedded in loose-fitting clothing by modeling secondary motion effects. MagShield (Shao et al. 2025) detects and corrects magnetic disturbances that corrupt orientation estimates in real-world environments. Group Inertial Poser (Xue et al. 2025) extends to multi-person tracking using inter-person UWB distances for relative positioning.

The field has converged on: (1) deep neural networks for learning motion priors from large datasets, (2) physics-based refinement for physical plausibility, (3) multi-modal fusion for enhanced accuracy, and (4) dynamic calibration for practical deployment. Global translation estimation, particularly in the vertical direction, remains difficult, with physics-based contact reasoning (Yi et al. 2025) providing the current best solution.

Sparse inertial motion-capture leaderboard. Method links to the chapter section, with citation attached. Year is conference/arXiv. #IMUs is the canonical sensor count (6 = standard wrists/shins/head/pelvis; 2 = consumer device pair; 3 = head+wrists). Output: K = local kinematic pose only, K+T = pose plus global translation, K+T+F = pose, translation, plus contact forces/torques, M = SMPL mesh. Latency end-to-end on the paper's reference hardware (Intel laptop + RTX 2080/3090/4090 class), in milliseconds. MPJPE is mean per-joint position error in millimetres on the paper's primary benchmark; "DIP" = DIP-IMU, "TC" = TotalCapture, "AIST" = AIST++, "UIP" = UIP-DB, "GIP" = GIP-DB. Jitter (root-MS jerk, m/s3^3) and foot-skate (cm/s during contact) are reported when the paper publishes them; " - " means unreported in primary configuration. Numbers reflect the primary published configuration and should be read with [fn:s0-imc-numbers-caveat] before cross-method ranking. Five cells are known to disagree with their sources and are left in place with the discrepancy recorded here rather than silently re-tabulated. (i) The PIP cell is the online DIP-IMU mesh error rather than MPJPE: Table 3 of (Yi et al. 2022) gives 5.04 cm positional error and 5.95 cm mesh error online, with absolute jitter 0.24, so the 0.16 entry matches no reported figure. (ii) The TransPose cell is the offline positional error (4.90 cm) while the PIP cell is online, so the two are not in the same setting and their ordering here inverts the within-paper comparison, in which PIP is ahead of TransPose. (iii) PIP publishes no ground-penetration rate at all, measuring plausibility by ZMP distance, so the foot-skate entry is unsupported. (iv) The BaroPoser cell is the DIP-IMU mesh error (10.36 cm) rather than the positional error (8.57 cm) reported in (Zhang et al. 2025), and its latency is derived from a 30 fps throughput rather than measured. (v) The Loose Inertial Poser row describes a 4-IMU upper-body system evaluated on its own loose-wear recordings, which reports nothing on DIP-IMU (Zuo et al. 2024); and the GIP cell could not be traced to any table of (Xue et al. 2025), whose real-data joint error is 9.94 cm. The DIP and TIP rows have no bibliography entry of their own: their figures are re-evaluations quoted by the citing papers, not values taken from the original publications.
Latency MPJPE Jitter Foot-skate
Method Year #IMUs Output (ms) (mm) (m/s3^3) (cm/s) Notes
DIP (Yi et al. 2021) 2018 6 K - 70 (DIP) 23.4 - bi-LSTM baseline; jitter unusable for graphics, no physics[fn:s0-flat-ground]
TransPose (Yi et al. 2021) 2021 6 K+T 94 49 (DIP) 1.19 - three-stage leaf-to-rotation cascade, offline setting; flat-ground[fn:s0-flat-ground] FK branch
PIP (Yi et al. 2022) 2022 6 K+T+F 16 59 (DIP) 0.16 <1% pen. dual-PD QP physics, online setting; plausibility reported as ZMP distance, not a penetration rate; μ\mu=0.5 fixed
TIP (Transformer Inertial Poser) 2022 6 K+T - 58 (TC) - - early transformer-only baseline; no physics, no calibration
PNP (Yi et al. 2024) 2024 6 K+T+F 16 49 (TC) - - fictitious-force compensator over PIP; biggest gain on fast motion
Loose Inertial Poser (Zuo et al. 2024) 2024 6 K+T - 65 (DIP) - - 4 jacket IMUs (both forearms, back, waist), upper-body pose; own loose-wear set, not DIP-IMU; SeMo-AE augmentation
Ultra Inertial Poser (UIP) (Armani et al. 2024) 2024 6+UWB K+T - 107 (UIP) 0.055 - inter-sensor UWB; SIP 24.1 deg; jitter -97% vs PIP; no abs. frame[fn:s0-no-absolute-anchor]
DiffCap (Pan et al. 2025) 2025 6+RGB K+T+M 320 31 (AIST) - - two-stage diffusion; MPJPE 60.5 on 3DPW-OCC vs PIP 97.8
TIC (Zuo et al. 2025) 2025 6 calib. <1 - - - drop-in calib. preprocessor; "put on and use"; no T-pose; orient. only[fn:s0-mag-yaw-drift]
MagShield (Shao et al. 2025) 2025 6 calib. 2 - - - yaw-only mag. correction[fn:s0-mag-yaw-drift]; pairs with PNP/DynaIP
BaroPoser (Zhang et al. 2025) 2025 2+baro K+T 33 104 (DIP) - - smartwatch+phone; barometer breaks flat-ground assumption
GlobalPose (Yi et al. 2025) 2025 6 K+T+F - 47 (TC) - - multi-contact (hands/knees) physics; 4.7% drift over 20-min sequences
Group Inertial Poser (GIP) (Xue et al. 2025) 2025 12+UWB K+T (NN=2) - 119 (GIP) - - first 2-person sparse-IMU; inter-person UWB anchors; N≥3N\geq3 unbenchmarked

The leaderboard reads in three method families, each defined by a different relaxation of an upstream assumption. The Yi-lab pose cascade (DIP →\to TransPose →\to PIP →\to PNP →\to GlobalPose) is the dominant inheritance line: each generation lifts one assumption while preserving the canonical 6-IMU wrists/shins/head/pelvis placement and the three-LSTM leaf-to-position-to-rotation backbone. TransPose reduced jitter to a level usable in graphics pipelines, roughly 19-fold below DIP; PIP added the dual-PD quadratic-program physics layer that makes the recovered motion satisfy an equation of motion and pushed end-to-end latency from 94 ms down to 16 ms; PNP fixed the inertial-frame assumption at the input stage, with biggest gains exactly where pelvis acceleration is largest (running, jumping); GlobalPose generalised contact to arbitrary joints, enabling stairs, sitting, and handstands without scene priors. The recurring failure mode across this entire line is the flat-ground assumption[fn:s0-flat-ground], inherited from TransPose's supporting-foot FK integrator, that GlobalPose only partially relaxes.

The fusion family (DiffCap, BaroPoser, UIP, GIP) breaks one of three IMU-only limitations by adding an exogenous modality. DiffCap adds a monocular camera and uses diffusion to absorb occlusion at the sequence level; its 17.4 mm MPJPE advantage on 3DPW-OCC over the previous best fusion method (60.5 against RobustCap's 77.9) is occlusion-specific and shrinks to 2.5 mm on clean TotalCapture (46.2 against 48.7) (Pan et al. 2025), so the modality matters more than the diffusion. BaroPoser adds barometric pressure from a smartwatch plus phone and is the first IMC method to estimate vertical translation on stairs without a learned height prior. Ultra Inertial Poser adds peer-to-peer UWB ranging between body-worn nodes and reduces jitter by roughly 97% via smoothly-varying distance constraints, while remaining anchor-free (and therefore unable to recover absolute world coordinates[fn:s0-no-absolute-anchor]). Group Inertial Poser extends the same UWB principle across persons and is the first synchronised two-person sparse-IMU system; N≥3N\geq3 compounding occlusion is unbenchmarked.

The calibration and robustness sidecar (TIC, MagShield, Loose Inertial Poser) is a pre-processing layer rather than a pose estimator: TIC drops T-pose calibration and operates as a transformer-based implicit calibrator over a sliding window; MagShield gates magnetometer corrections under detected magnetic disturbance and is yaw-only by construction (since gravity already constrains roll and pitch[fn:s0-mag-yaw-drift]); Loose Inertial Poser models secondary-motion offsets between garment and body so that the static RBSR_{BS} calibration assumption can be relaxed for everyday clothing. These three are complementary (TIC corrects calibration drift, MagShield corrects magnetic interference, LIP corrects garment offset) and are commonly stacked as parallel preprocessors before a downstream PIP/PNP/DynaIP pose estimator. Cross-paper number comparison is unsafe[fn:s0-imc-numbers-caveat]; within-paper ablations are the reliable signal.

References

Armani, Rayan, Changlin Qian, Jiaxi Jiang, and Christian Holz. 2024. Ultra Inertial Poser: Scalable Motion Capture and Tracking from Sparse Inertial Sensors and Ultra-Wideband Ranging. arXiv. https://doi.org/10.48550/arXiv.2404.19541.
Pan, Shaohua, Xinyu Yi, Yan Zhou, et al. 2025. DiffCap: Diffusion-Based Real-Time Human Motion Capture Using Sparse IMUs and a Monocular Camera. arXiv. https://doi.org/10.48550/arXiv.2508.06139.
Shao, Yunzhe, Xinyu Yi, Lu Yin, Shihui Guo, Junhai Yong, and Feng Xu. 2025. MagShield: Towards Better Robustness in Sparse Inertial Motion Capture Under Magnetic Disturbances. arXiv. https://doi.org/10.48550/arXiv.2506.22907.
Xue, Ying, Jiaxi Jiang, Rayan Armani, Dominik Hollidt, Yi-Chi Liao, and Christian Holz. 2025. Group Inertial Poser: Multi-Person Pose and Global Translation from Sparse Inertial Sensors and Ultra-Wideband Ranging. arXiv. https://doi.org/10.48550/arXiv.2510.21654.
Yi, Xinyu, Shaohua Pan, and Feng Xu. 2025. Improving Global Motion Estimation in Sparse IMU-Based Motion Capture with Physics. arXiv. https://doi.org/10.48550/arXiv.2505.05010.
Yi, Xinyu, Yuxiao Zhou, Marc Habermann, et al. 2022. Physical Inertial Poser (PIP): Physics-Aware Real-Time Human Motion Tracking from Sparse Inertial Sensors. arXiv. https://doi.org/10.48550/arXiv.2203.08528.
Yi, Xinyu, Yuxiao Zhou, and Feng Xu. 2021. TransPose: Real-Time 3D Human Translation and Pose Estimation with Six Inertial Sensors. arXiv. https://doi.org/10.48550/arXiv.2105.04605.
Yi, Xinyu, Yuxiao Zhou, and Feng Xu. 2024. Physical Non-Inertial Poser (PNP): Modeling Non-Inertial Effects in Sparse-Inertial Human Motion Capture. arXiv. https://doi.org/10.48550/arXiv.2404.19619.
Zhang, Libo, Xinyu Yi, and Feng Xu. 2025. "BaroPoser: Real-Time Human Motion Tracking from IMUs and Barometers in Everyday Devices." Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology, September, 1-9. https://doi.org/10.1145/3746059.3747731.
Zuo, Chengxu, Jiawei Huang, Xiao Jiang, et al. 2025. "Transformer IMU Calibrator: Dynamic On-Body IMU Calibration for Inertial Motion Capture." ACM Transactions on Graphics 44 (4): 1-14. https://doi.org/10.1145/3730937.
Zuo, Chengxu, Yiming Wang, Lishuang Zhan, et al. 2024. "Loose Inertial Poser: Motion Capture with IMU-Attached Loose-Wear Jacket." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Seattle, WA, USA), June, 2209-19. https://doi.org/10.1109/CVPR52733.2024.00215.