Skeleton-Based Posture Classification to Promote Safer Walker-Assisted Gait in Older Adults

Technical Summary: Skeleton-Based Posture Classification for Safer Walker-Assisted Gait

Problem Statement
The aging global population faces significant health challenges, including declines in balance, muscle mass, and mobility, which drastically increase the risk of falls. While assistive devices like walkers are standard for mitigating these risks, their misuse—specifically incorrect posture or improper handling—can paradoxically lead to falls. Current posture classification solutions often rely on full-body datasets requiring substantial processing power or wearable markers, which can be intrusive, costly, and difficult for older adults to maintain. There is a need for a non-intrusive, real-time system capable of monitoring user posture using on-board cameras on assistive devices to detect hazardous postures and prevent falls without requiring the user to wear additional sensors.

Methodology
The study proposes and compares five distinct modeling approaches for classifying walker usage, initial position (sitting vs. standing), and specific postures using skeleton-based data. The system utilizes a Socially Assistive Walker (SAW) equipped with a USB camera, LiDARs, force-sensing resistors (FSRs), and an Inertial Measurement Unit (IMU).

1. Data Acquisition and Preprocessing:

   • Data Collection: Data was gathered from 21 healthy participants (ages 21–48) performing predefined exercises (arm raises, torso bends, sitting, standing) under controlled conditions.
   • Pose Landmark Detection: Google MediaPipe's Pose Detection framework (BlazePose) was employed to extract 33 three-dimensional body landmarks in real-time. This approach avoids the need for raw image processing, transmitting only text-based coordinates for efficiency.
   • Feature Extraction: Raw landmark coordinates were pre-processed to extract 48 geometric features, including distances and angles between joints.

2. Model Architectures:

   • Convolutional Neural Networks (CNN): Two architectures were tested: a 4-layer CNN (tested with both raw landmarks and engineered features) and a 6-layer CNN (raw landmarks only). These models utilize convolution, batch normalization, max pooling, and dropout layers.
   • Encoder-Decoder CNN (EDCNN): An architecture designed for dimensionality reduction and reconstruction, featuring an encoder with three convolutional layers and a decoder with transposed convolutional layers, trained jointly for classification.
   • Support Vector Machine (SVM): A supervised algorithm used to identify optimal hyperplanes for separating classes.
   • Extreme Gradient Boosting (XGBoost): An ensemble learning method using decision trees, combined with a MultiOutputClassifier to predict multiple variables simultaneously.
   • Geometric Approach: A rule-based, non-learning method that establishes a baseline pose and classifies deviations using Euclidean distances and empirically determined thresholds, augmented by FSR data for hand-lifting detection.

3. Classification Tasks:
   The models were trained to predict three outputs:

   • Choice of Walker: Binary classification (walker used vs. not used).
   • Initial Position: Binary classification (sitting vs. standing).
   • Posture Classification: Multi-class classification (17 specific postures for supervised models; 8 postures for the geometric approach).

Key Results
The study evaluated performance using accuracy, precision, F1 score, sensitivity, and specificity on both training/validation sets and unseen test data.

• Binary Classification (Walker Choice & Initial Position):

  • XGBoost emerged as the top performer, achieving near-perfect training accuracy (100% for walker choice, 99.9% for initial position) and maintaining high prediction accuracy (99.84% and 99.69%, respectively).
  • Deep Learning Models: The 4-layer CNN, 6-layer CNN, and EDCNN also demonstrated strong performance in binary tasks, with training accuracies generally exceeding 98%.
  • SVM showed variable performance, achieving 97.4% accuracy for walker choice but lower results for initial position (90.3%).

• Multi-Class Posture Classification:

  • Challenges: All models faced significant difficulty with the 17-class posture task compared to binary tasks.
  • XGBoost: Achieved 99.24% training accuracy but dropped to 74.1% prediction accuracy for the 17-class task. When simplified to a single-output risk classification (standing, sitting, or bad posture), it achieved 85.0% prediction accuracy.
  • Deep Learning: The 4-layer CNN achieved the highest prediction accuracy for the 17-class task among deep learning models at 76.0%. The EDCNN and 6-layer CNN showed lower prediction accuracies (66.0% and 75.7%, respectively).
  • Geometric Approach: In a specific test with 8 postures, this approach achieved an overall accuracy of 89.87%. It performed exceptionally well for static poses (Standing: 98.15%) and hand-lifting actions (98%+), though performance dropped for fall simulations (e.g., Fall Backward: 60.02%).

• Comparative Metrics:

  • For binary tasks, XGBoost and the Geometric approach generally yielded the highest F1 scores.
  • Deep learning models showed competitive performance but struggled more with the complexity of multi-class classification, particularly as architectural depth increased (e.g., 6-layer CNN).

Significance and Claims
The paper claims that machine learning, particularly XGBoost and geometric approaches, holds significant potential for enhancing human-robot interaction in smart walkers for fall prevention. Key contributions include:

1. Feasibility of On-Board Processing: The study demonstrates that using skeleton landmarks (text data) rather than raw images allows for efficient processing, making real-time monitoring feasible on resource-constrained devices like the Raspberry Pi 4.
2. Superiority of XGBoost for Binary Tasks: XGBoost is identified as the most effective model for determining critical binary states (walker usage, sitting vs. standing), which are paramount for triggering safety mechanisms like brakes during sit-to-stand transfers.
3. Risk Factor Detection: While multi-class posture classification remains challenging, the ability to detect "bad postures" or deviations from standing (e.g., leaning, lifting hands) with reasonable accuracy (83% F1 score for XGBoost single-output) offers a viable method for early fall risk detection.
4. Privacy and Non-Intrusiveness: By utilizing marker-less camera systems mounted on the walker, the approach avoids the privacy concerns and usability barriers associated with wearable sensors.

Limitations Acknowledged
The authors note that the dataset was relatively small (21 participants), limiting generalization. Furthermore, multi-class posture classification accuracy remains lower than binary tasks, suggesting a need for more sophisticated architectures. The study also highlights that computational requirements for deep learning may hinder deployment on single-board computers and that latency and power consumption were not assessed. Future work is directed toward real-world validation and integrating additional sensor modalities.