JMIR Publications

ABSTRACT

Background:

Artificial Intelligence (AI)-driven pose estimation offers a scalable and cost-effective solution to track exercises in mobile health (mHealth) applications. However, occlusion, influenced by camera angle and distance, can reduce detection accuracy and repetition counting precision. The influence of smartphone positioning on these performance metrics remains underexplored in controlled studies.

Objective:

To examine how smartphone camera angle (front, side, and diagonal) and distance (90 cm, 180 cm, 200 cm, and 360 cm) affect detection performance and repetition counting accuracy during push-ups and squats using AI-based pose estimation.

Methods:

In this cross-sectional, within-subject study, 44 healthy university students (nine female [20.5%]; mean age 20.3 ± 0.4 years; mean BMI 23.2 ± 0.6 kg/m²) were assigned to perform either squats or push-ups. Each participant completed their assigned exercise across 12 predefined smartphone camera configurations, yielding ~264 squat trials (n=22) and 264 push-up trials (n=22). Each trial averaged five repetitions, totaling ~1,320 repetitions per exercise. Pose estimation performance was assessed using binary classification accuracy, detection rate, and mean absolute error (MAE) for repetition counting. Generalized linear mixed-effects models (GLMMs) evaluated classification odds, linear mixed-effects models (LMMs) analyzed MAE, and Tukey-adjusted post hoc tests followed significant effects.

Results:

The mean detection rate was 61.1% (SD 48.8%) for push-ups and 61.5% (SD 48.7%) for squats, with MAEs of 1.08 (SD 1.78) and 1.11 (SD 1.82) repetitions, respectively. Push-ups were most accurately detected from diagonal views at 90–180 cm (up to 85.7% detection; MAE = 0.28) and least accurately from the front at 360 cm (20.0%; MAE = 2.70). Squats performed best from a diagonal view at 200 cm (95.5%; MAE = 0.05) and worst from the side at 90 cm (0.0%; MAE = 5.00). GLMMs showed that for push-ups, the front 90 cm and diagonal 360 cm views significantly reduced classification odds compared to the side 90 cm view (p = 0.032 and p = 0.040, respectively), whereas for squats, diagonal and front views significantly outperformed side views across all distances (p < .001). Post hoc tests confirmed that for push-ups, diagonal close/mid-range views had significantly lower MAEs than far front views, and for squats, diagonal and front views at 180-200 cm achieved the highest accuracy and lowest MAEs (p < .05).

Conclusions:

AI-based pose estimation effectiveness for exercise tracking is significantly affected by smartphone positioning. Diagonal and frontal views at mid-range distances (180–200 cm) provided the highest detection accuracy and counting precision. These findings offer actionable guidance for developers, clinicians, coaches, and users optimizing mHealth exercise monitoring.