Class-Agnostic Repetitive Action Counting Using Wearable Devices.

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Machine Learning

Background:

• Repetitive human action counting is crucial for various applications.
• Existing methods often require extensive labeled data or are limited to specific action classes.
• Class-agnostic counting using time-series sensor data remains a significant challenge.

Purpose of the Study:

• To introduce Class-agnostic Repetitive action Counting (CaRaCount), a novel few-shot learning approach.
• To enable the counting of any repetitive human action using wearable device time-series data.
• To evaluate the performance and usability of CaRaCount in diverse, real-world scenarios.

Main Methods:

• Developed CaRaCount, a few-shot, class-agnostic method for action counting.
• Collected a large-scale time-series dataset of repetitive human actions using smartwatch data.
• Validated CaRaCount on the newly collected dataset and three existing datasets (Crossfit, Recofit, MM-Fit).

Main Results:

• CaRaCount achieved low error rates in counting repetitive actions across multiple datasets.
• The method demonstrated superior performance compared to existing baselines and state-of-the-art techniques.
• A user experience study confirmed the usability of the real-time CaRaCount implementation.

Conclusions:

• CaRaCount is an effective and efficient approach for real-world repetitive action counting.
• The few-shot, class-agnostic nature of CaRaCount significantly advances the field.
• The approach shows strong potential for deployment outside controlled laboratory settings.