A Deep Learning Approach for Foot Trajectory Estimation in Gait Analysis Using Inertial Sensors

Area of Science:

• Biomedical Engineering
• Gerontology
• Machine Learning

Background:

• Gait performance is a key indicator of motor and cognitive health in older adults.
• Instrumented gait analysis using inertial sensors offers a portable alternative to laboratory assessments.
• Estimating gait parameters from inertial sensors requires overcoming cumulative integration errors.

Purpose of the Study:

• To develop a deep recurrent neural network for accurate heel and toe trajectory estimation.
• To propose a coordinate frame transformation to simplify stride trajectory analysis.
• To estimate a comprehensive set of spatiotemporal gait parameters using the novel approach.

Main Methods:

• A deep recurrent neural network was employed to predict foot trajectories.
• A novel coordinate frame transformation was introduced for stride trajectories.
• Validation was performed using optical motion capture data from young adults.

Main Results:

• Heel and toe trajectories were predicted with low errors, closely matching reference data.
• Estimated spatiotemporal gait parameters showed good agreement with reference values, especially when excluding turning strides.
• The method demonstrated robustness to imperfect sensor-foot alignment.

Conclusions:

• The proposed deep recurrent neural network approach accurately estimates foot trajectories and gait parameters.
• This method provides a reliable and robust tool for instrumented gait analysis.
• The findings support the use of inertial sensors for comprehensive gait assessment in various conditions.