A Sequential-Data Approach for Human Gait Parameter Estimation and Surgical Event Inference

Abstract

Human gait analysis is performed in clinical research studies by estimating relevant gait parameters (GPs). Conventional approaches for gait analysis involve expensive wearable sensors for three-dimensional motion capture requiring specialized knowledge while possessing limited data. Video-based pose estimation techniques such as Openpose provide the key points data using a two-dimensional video, which is easy to record without the need of specialized equipment or personnel. This research presents estimation of human gait parameters using sequential key points data extracted from 2-D videos. Key points are extracted using open pose and sequential data is furnished, deep learning network developed, followed by gait parameter (GP) estimation including Speed, Cadence, Gait deviation index and Knee flex. After development, rigorous analysis of 24 architectures of recurrent neural network (RNN) is performed and the best two models are chosen. For performance evaluation, correlation between sensor-based recorded true and video-based estimated gait parameters is computed, as well as training vs. validation loss and computational complexity. Furthermore, to strengthen the results, error analysis and demographic analysis are presented. The experimental results demonstrate that the proposed RNN-based architectures outperform the state-of-the-art convolutional neural network (CNN) and machine learning based models. The results for RNN and CNN are also evaluated on gait identification dataset which states the superior RNN performance as well. Moreover, the surgical event diagnostic inference is provided using the estimated gait parameters.