BlazePose-Seq2Seq: Leveraging Regular RGB Cameras for Robust Gait Assessment

Area of Science:

• Biomedical Engineering
• Computer Science
• Kinetics

Background:

• Smartphone-based pose estimation offers a cost-effective alternative to traditional gait analysis methods.
• Current algorithms like BlazePose show promise but lack clinical-level accuracy and repeatability.
• Sequence-to-sequence (Seq2seq) architectures present a novel approach for improving kinematic predictions.

Purpose of the Study:

• To enhance the BlazePose algorithm using a Seq2seq autoencoder for more accurate human gait analysis.
• To evaluate the effectiveness of deep learning models in improving smartphone-based pose estimation for gait kinematics.
• To establish a reliable and accurate method for remote gait assessment.

Main Methods:

• A novel deep learning model combining BlazePose with Seq2seq autoencoders was developed.
• Synchronized motion capture data from RGB and Vicon cameras were collected at three walking speeds.
• Performance was evaluated by comparing BlazePose alone against integrated models (1D-LSTM, GRU, LSTM) using mean absolute error.

Main Results:

• The integrated BlazePose and Seq2seq models significantly reduced mean absolute errors in joint angle prediction.
• Average errors decreased from 13.4° to 5.3° (fast gait), 16.3° to 7.5° (normal gait), and 15.5° to 7.5° (slow gait) at the left ankle.
• The enhanced model demonstrated improved accuracy and repeatability for gait analysis.

Conclusions:

• Incorporating Seq2seq architectures into smartphone-based pose estimation substantially improves gait analysis accuracy.
• This approach offers a viable solution for remote, clinical-grade gait assessment.
• The study validates the use of deep learning for advancing virtual motion analysis.