Viewpoint-invariant exercise repetition counting.

Area of Science:

• Computer Vision
• Biomedical Engineering
• Human Motion Analysis

Background:

• Accurate repetition counting is crucial for effective physical rehabilitation and exercise training.
• Existing vision-based methods struggle with concurrent motions and diverse camera viewpoints.
• There is a need for robust repetition counting systems that are independent of viewing angle and motion complexity.

Purpose of the Study:

• To develop and validate a novel method for counting human exercise repetitions using spectrogram analysis of pose estimation.
• To demonstrate the method's effectiveness in handling concurrent motions and different view angles.
• To assess the method's performance on public datasets and real-world collected data.

Main Methods:

• Analyzed the spectrogram of pose estimation cosine similarity for repetition counting.
• Validated the method on the UI-PRMD and MM-fit public datasets.
• Collected and tested the method on exercise videos from 11 adults, including concurrent motions and various camera locations.

Main Results:

• Achieved a mean absolute error (MAE) of 0.06 and off-by-one accuracy (OBOA) of 0.94 on the MM-fit dataset.
• Achieved an MAE of 0.06 and OBOA of 0.95 on the UI-PRMD dataset.
• Demonstrated an overall MAE of 0.07 and OBOA of 0.91 across 57 skeleton time-series videos with varied camera angles and concurrent motions.

Conclusions:

• The proposed method provides view-angle and motion-agnostic concurrent motion counting for human exercises.
• This technique shows significant potential for large-scale remote rehabilitation and exercise training applications using a single camera.
• The method offers a reliable and versatile solution for automated exercise analysis.