Faster Deep Inertial Pose Estimation with Six Inertial Sensors

Area of Science:

• Computer Vision
• Human Pose Estimation
• Wearable Technology

Background:

• Traditional vision-based pose estimation faces challenges like occlusion and high deployment costs.
• Inertial Measurement Units (IMUs) offer a viable alternative but require robust processing for accurate pose prediction.

Purpose of the Study:

• To introduce a novel, efficient, and accurate full-body pose estimation method using minimal inertial sensors.
• To address computational efficiency and data noise issues in IMU-based pose tracking.

Main Methods:

• Utilized a novel Synchronized Recurrent Unit (SRU) network architecture, replacing traditional bidirectional RNNs for reduced computational load.
• Developed a model that achieves state-of-the-art results without requiring explicit joint position supervision.
• Implemented SmoothLoss to mitigate the impact of noisy inertial sensor data on pose estimation accuracy.

Main Results:

• The proposed Faster Deep Inertial Poser (FDIP) model achieves online inference speeds of 90 FPS on a CPU.
• Demonstrated a reduction in error impact by over 10% compared to previous methods.
• Achieved a 250% increase in inference speed over existing state-of-the-art techniques.

Conclusions:

• The FDIP model offers a computationally efficient and accurate solution for full-body pose estimation using inertial sensors.
• This method provides a practical alternative to vision-based systems, overcoming limitations like occlusion.
• The approach significantly enhances both the speed and accuracy of IMU-based human pose tracking.