• Pytorh GPU implementation of the FDRNet - a new deep learning autoencoder architecture for multivariate long sequence generation
• Basic Autoencoder Architecture for Long-Sequence Multivariate Time-Series Learning

Gait analysis is the systematic study of human locomotion using external visual and sensory observations. Data extraction from such observations enables the quantitative assessment of human motion skills that can lead to feedback discovery and physical ability evaluation. Current research methods in human gait analysis cover a substantial portion on gait- phase identification, locomotion recognition and gait trajectory prediction. Although recognizing and predicting human phase-gait cycles is important for un- derstanding and analyzing physical motor ability, it is insufficient to provide objective gait rate feedback and motor function recovery. In this work we introduce and implement the Feedback Recovery Network (FDRNet). FDRNet is a novel deep neural network architecture able to provide objective personalized quan- titative gait feedback discovery and correction on multivariate time-series data. We train FDRNet using an online available dataset containing 1020 multivariate gait signals from 230 subjects undergoing a fixed protocol: standing still, walking 10 m, turning around, walking back and stopping. The measured population is composed of healthy subjects as well as patients with neurological or orthopedic disorders. The FDRNet holds promise for automation in personalized rehabilitation and considers the potential of creating new technology in the area of physical ability performance assessments.

• All source code is in source.
• Train using the main.py file.
• Get model predictions using the predict.py file.

• Code is the documentation of itself

• Use python3 main.py to train your feedback discovery network.
• A full report of the pipeline is given in Project-Report.pdf.

The pipeline is demonstrated below.

#- Training Curves.

#| Loss Curve #| --- | #|

• Feedback Discovery on a sample of 16-channel time series per 1042 time-step signal.

Left Foot Vertical Acceleration [LAV]

Petros Apostolou and Vassilios Morellas, Gait Feedback and Correction Generation Using Multivariate Sequence Learning, Computer Science and Engineering Department, University of Minnesota, 2022.

• apost035@umn.edu, trs.apostolou@gmail.com