Wearable Spine Tracker vs. Video-Based Pose Estimation for Human Activity Recognition.

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Rehabilitation Technology

Background:

• Activities of daily living (ADLs) recognition is crucial for healthcare and assistive technologies.
• Wearable sensors and computer vision offer distinct approaches for ADL monitoring.
• A direct comparison of these systems using standardized protocols is needed.

Purpose of the Study:

• To comparatively evaluate the performance of a wearable spine tracker (FlexTail) and a camera-based pose estimation model for ADL detection.
• To assess various time series classification algorithms for ADL recognition accuracy.
• To investigate the effect of hierarchical activity grouping on classification performance.

Main Methods:

• Developed a protocol for simultaneous data acquisition using FlexTail and camera systems.
• Recorded eleven distinct ADLs encompassing general movement, household chores, and food handling.
• Applied and compared state-of-the-art time series classification algorithms, including Random Dilated Shapelet Transform (RDST) and QUANT classifier.

Main Results:

• Both FlexTail and camera systems achieved high average F1 scores of 0.90 using a 1-second window.
• RDST classifier was optimal for FlexTail data, while QUANT classifier performed best for camera data.
• Hierarchical activity grouping showed inconsistent benefits across different activities.

Conclusions:

• Both wearable spine trackers and camera-based systems are effective for recognizing ADLs.
• FlexTail demonstrates superior performance in detecting postural transitions (e.g., sitting, standing).
• Camera-based systems excel at recognizing activities involving fine motor skills and arm movements.