ai push up counter: on-device real-time counting

On-device AI pipeline for real-time push-up counting

Core components

This ai push up counter runs entirely on-device and counts reps in real time without sending footage off-device.

• Pose estimation: a lightweight model runs on-device from the camera feed, extracting keypoints to map push-up form.
• On-device inference: all nets run locally with quantized, mobile-optimized architectures to minimize memory use.
• Motion analysis and counting: temporal signals from keypoints compute angles and chest distance; the system detects down and up phases to count reps.
• Real-time counting: continuous loop delivering immediate feedback; typical performance on mid-range devices sits in the 15–30 FPS range with low latency.
• Privacy and no wearables: no data leaves the device; no extra sensors needed.
• Cross-platform: a single core pipeline runs on iOS and Android; optimized path adapts to each platform’s camera stack.

Why on-device inference matters

• Privacy stays local to the device.
• Low latency enables real-time coaching cues.
• Offline use and a small memory footprint support low-end devices.

Calibration, robustness, and real-world accuracy

Calibration workflow

• Start with a quick framing check to ensure the camera sees the torso and arms clearly, with the subject centered and within a safe zone.
• Run a brief auto-check on pose visibility and lighting, then prompt the user to perform 2–3 push-ups in different styles (standard, wide, incline) to capture a range of motion.
• Build a per-user motion template that adjusts for body types and limb lengths, so elbow and shoulder angles map consistently to counts.
• Prompt a quick camera-angle tune, guiding the user to reduce extreme tilts or distance changes that blur keypoints.
• Normalize lighting locally and store calibration data on-device as a profile that the real-time counter will reference.

Handling edge cases

• Occlusions are handled with temporal smoothing and partial keypoint confidence; counting pauses gracefully if data drops below a threshold and recovers when reliable.
• If multiple people enter frame, the system tracks the primary user by proximity and initial seed, ignoring others for counting.
• Background clutter is suppressed with lightweight filtering that ignores non-human motion and uses movement context to maintain robustness.

Privacy-first UX and expanding to other exercises

Privacy considerations

• All push-up counting and related metrics are computed on-device, with no data transmitted by default.
• Privacy policy presented during onboarding; you can revisit it later.
• Local data handling: counts, form cues, and progress are stored locally and can be deleted easily; none are sent to cloud unless user explicitly enables sharing.
• Offline operation: the app works without internet; you can use it anywhere.
• Optional offline model updates: you can download model updates as a package and apply them locally; no automatic network calls.
• Clear user expectations: we won't surprise you with background uploads; you control data and updates.

Extending to squats and pull-ups

• A shared pipeline will power counting for new exercises (squats, pull-ups) using the same on-device architecture.
• Cross-platform compatibility: iOS and Android share the same model and UI patterns for consistency.
• Roadmap: additional exercises roll out as optional modules that update the local model package; you’ll see clear in-app prompts and an updated privacy policy if needed.