Time Coherent Full-Body Poses Estimated Using Only Five Inertial Sensors: Deep versus Shallow Learning

Area of Science:

• Biomechanical Engineering
• Computer Vision
• Machine Learning

Background:

• Traditional full-body motion capture is hindered by lengthy setup times and obtrusive sensors on each body segment.
• Reducing sensor count via deep learning or offline methods necessitates extensive datasets and significant computational power.

Purpose of the Study:

• To evaluate the performance of a shallow learning approach against deep learning for full-body pose estimation using only five inertial sensors.
• To investigate the impact of incorporating past/future inertial sensor data and acceleration input on pose estimation accuracy and efficiency.

Main Methods:

• A shallow neural network was developed, utilizing a stacked input vector comprising past and future inertial sensor information.
• Shallow and deep learning models were compared using identical input vector configurations, including the evaluation of acceleration data.
• Performance was assessed based on pose estimation accuracy, jerk errors, and system delay.

Main Results:

• The shallow learning approach achieved comparable accuracy (~6 cm) to the deep learning approach (~7 cm) in full-body pose estimation.
• Deep learning exhibited lower jerk errors, potentially due to its explicit recurrent modeling capabilities.
• The shallow learning approach demonstrated a significantly smaller system delay (72 ms) compared to deep learning (117 ms).

Conclusions:

• Shallow learning offers a viable and efficient alternative for full-body motion capture using minimal inertial sensors, achieving high accuracy.
• The reduced delay in the shallow learning method makes it advantageous for real-time applications where latency is critical.
• Further research could explore hybrid models to leverage the strengths of both shallow and deep learning approaches.