Human-centered design-based lightweight wearable IMU human pose estimation

Area of Science:

• Computer Vision
• Machine Learning
• Wearable Technology

Background:

• Human pose estimation is crucial for human-computer interaction.
• Existing methods often require significant computational resources, limiting on-device deployment.
• Wearable Inertial Measurement Unit (IMU) based systems offer a privacy-preserving alternative.

Purpose of the Study:

• To develop a highly efficient human pose estimation framework for on-device deployment on wearables.
• To achieve a balance between high accuracy and low latency for real-time applications.
• To leverage knowledge distillation and structural re-parameterization for model optimization.

Main Methods:

• A Transformer-based teacher model learns spatio-temporal representations.
• An involution-based student model utilizes input-adaptive operators.
• Structural re-parameterization collapses the training graph for efficient inference.
• Knowledge distillation transfers learning from the teacher to the student model.

Main Results:

• The proposed framework achieves near state-of-the-art accuracy (81 mm MPJPE on DIP-IMU, 94 mm on IMUPoser).
• Sub-millisecond latency (0.012 ms on DIP-IMU, 0.011 ms on IMUPoser) is attained.
• Significant speedups (one to two orders of magnitude) over Transformer baselines are demonstrated.
• The model shows robustness and cross-subject generalization.

Conclusions:

• The developed framework is hardware-friendly and suitable for low-power wearables.
• It enables efficient, real-time human pose estimation on edge devices.
• The approach effectively decouples training expressiveness from inference efficiency.