Evaluation of Smartphone Camera Positioning on Artificial Intelligence Pose Estimation Accuracy for Exercise Detection: Observational Study.

Overall mean detection rates were approximately 61% for both push-ups and squats, with notable error in repetition counting.

• Mean detection rate was 61.1% (SD 48.8%) for push-ups and 61.5% (SD 48.7%) for squats.
• Mean absolute error (MAE) for repetition counting was 1.08 (SD 1.78) repetitions for push-ups and 1.11 (SD 1.82) repetitions for squats.
• Study involved 44 healthy university students (mean age 20.3 y, SD 0.4 y; mean BMI 23.2, SD 0.6 kg/m²), with 22 participants per exercise.
• Each participant completed their assigned exercise across 12 predefined smartphone camera configurations, yielding approximately 264 trials per exercise and approximately 1320 repetitions per exercise.

Push-up detection accuracy and repetition counting precision varied substantially by camera angle and distance.

• Push-ups were most accurately detected from diagonal views at 90 to 180 cm, achieving up to 85.7% detection and MAE of 0.28 repetitions.
• Push-ups were least accurately detected from the front view at 360 cm, with only 20% detection and MAE of 2.70 repetitions.
• Generalized linear mixed models showed that front 90 cm and diagonal 360 cm views significantly reduced classification odds compared to the side 90 cm view (P=.03 and P=.04, respectively).
• Post hoc tests confirmed that diagonal close or mid-range views had significantly lower MAEs than far front views (P<.05).

Squat detection accuracy and repetition counting precision also varied substantially by camera angle and distance, with the worst performance from the side view at close range.

• Squats performed best from a diagonal view at 200 cm, achieving 95.5% detection and MAE of 0.05 repetitions.
• Squats performed worst from the side view at 90 cm, with 0% detection and MAE of 5 repetitions.
• For squats, diagonal and front views significantly outperformed side views across all distances (P<.001) in generalized linear mixed models.
• Post hoc tests confirmed that diagonal and front views at 180 to 200 cm achieved the highest accuracy and lowest MAEs for squats (P<.05).

Camera angle had a differential effect on pose estimation performance depending on exercise type, with side views performing relatively better for push-ups than for squats.

• For squats, side views at all distances were significantly outperformed by diagonal and front views (P<.001), indicating the side view is particularly unsuitable for squat detection.
• For push-ups, the side 90 cm view served as the reference condition against which front and diagonal views were compared, suggesting side views were among the better-performing configurations for push-ups.
• The front 90 cm view significantly reduced classification odds for push-ups compared to the side 90 cm reference (P=.03).
• The diagonal 360 cm view also significantly reduced classification odds for push-ups compared to the side 90 cm reference (P=.04).

Occlusion influenced by camera angle and distance was identified as a key factor reducing AI pose estimation detection accuracy and repetition counting precision.

• The study examined three camera angles (front, side, and diagonal) and four distances (90 cm, 180 cm, 200 cm, and 360 cm) across two exercises (push-ups and squats).
• The study used a cross-sectional, within-subject design where each participant completed all 12 camera configurations for their assigned exercise.
• Performance was assessed using binary classification accuracy, detection rate, and MAE for repetition counting.
• Generalized linear mixed-effects models evaluated classification odds, linear mixed-effects models analyzed MAE, and Tukey-adjusted post hoc tests followed significant effects.

Mid-range distances of 180 to 200 cm combined with diagonal or frontal camera angles provided optimal conditions for AI-based exercise detection across both push-ups and squats.

• Diagonal views at 90 to 180 cm achieved up to 85.7% detection for push-ups with MAE of 0.28.
• Diagonal views at 200 cm achieved 95.5% detection for squats with MAE of 0.05.
• The authors concluded that 'diagonal and frontal views at mid-range distances (180-200 cm) provided the highest detection accuracy and counting precision.'
• The findings were described as offering 'actionable guidance for developers, clinicians, coaches, and users optimizing mobile health exercise monitoring.'