A novel multi-level 3D pose estimation framework for gait detection of Parkinson's disease using monocular video

Abstract

Introduction: Parkinson's disease (PD) is characterized by muscle stiffness, bradykinesia, and balance disorders, significantly impairing the quality of life for affected patients. While motion pose estimation and gait analysis can aid in early diagnosis and timely intervention, clinical practice currently lacks objective and accurate tools for gait analysis.

Methods: This study proposes a multi-level 3D pose estimation framework for PD patients, integrating monocular video with Transformer and Graph Convolutional Network (GCN) techniques. Gait temporal and spatial parameters were extracted and verified for 59 healthy elderly and PD patients, and an early prediction model for PD patients was established.

Results: The repeatability of the gait parameters showed strong consistency, with most of the estimated parameters yielding an Intraclass Correlation Coefficient (ICC) greater than 0.70. Furthermore, these parameters exhibited a high correlation with VICON and ATMI results (r > 0.80). The classification model based on the extracted parameter features, using a Random Forest (RF) classifier, achieved an accuracy of 93.3%.

Conclusion: The proposed 3D pose estimation method demonstrates high reliability and effectiveness in providing accurate 3D human pose parameters, with strong potential for early prediction of PD.

Significance: This markerless method offers significant advantages in terms of low cost, portability, and ease of use, positioning it as a promising tool for monitoring and screening PD patients in clinical settings.

Keywords: 3D pose estimation; Parkinson’s disease (PD); gait detection; graph convolutional network (GCN); monocular video.

FIGURE 1

The flowchart and objectives of this research.

FIGURE 2

Experimental setup, (A) 3D plantar pressure measurement system setup, (B) depth camera position schematic, (C) VICON experimental configuration.

FIGURE 3

The flowchart of the method includes extracting 2D key points, Extracting 3D key points, feature extraction and data analysis. (A) Extract 2D key points. (B) Extract 3D key points (C) Feature extraction and analysis (i) Hourglass Tokenizer (HoT) Transformer Blocks. (ii) Global-local Adaptive Graph Convolutional Network (GLA-GCN).

FIGURE 4

Temporal feature extraction diagram. (A) HoT + Transformer. (B) Global-local Adaptive GCN. (C) Transformer + GCN.

FIGURE 5

Analysis of difference results. Box plot of the difference analysis of temporal features and spatial features extracted from video data among healthy people (yellow), carly PD (bluc), and mid-stage PD (purple) groups, represents p-value < 0.05, ** represents p-value < 0.01.

FIGURE 6

ROC curve and AUC calculation results of each classifier.