SmartForm

SmartForm is an Android fitness app that performs on-device pose detection and hand tracking to help users train with better form, count reps, and control the experience hands-free with gestures. It is built fully on-device with CameraX, ML Kit, MediaPipe, and Jetpack Compose.

It currently supports guided tracking for curls, squats, and push-ups, with posture-aware rep counting, per-exercise calibration, rep-quality scoring, and a graph-driven session summary — all wrapped in a dark, athletic "neon" UI.

Download

Grab the APK for your device from the latest release:

Variant	Devices	Size
arm64-v8a	Recommended — all modern phones (2017+)	~45 MB
armeabi-v7a	Older 32-bit devices	~36 MB
universal	Works on any device	~130 MB

Requires Android 8.0 (API 26)+. Also available on Uptodown:

Screenshots

Onboarding	Live workout HUD	Session summary	Session breakdown

Overview

SmartForm combines three capabilities in a single camera pipeline:

• Real-time body pose detection for exercise analysis
• Real-time hand landmark tracking for gesture input
• Live rep counting and form evaluation for supported exercises

The experience is designed around a hands-free workout loop:

1. Grant camera access
2. Choose an exercise mode (gesture or on-screen)
3. Start a session (pinch gesture or on-screen button)
4. Perform reps while SmartForm evaluates posture and rep quality
5. End the session and review the graph-based summary

Features

Real-time pose tracking

• Full-body pose detection using ML Kit Pose Detection
• Live skeleton overlay aligned to the camera preview
• Continuous, backpressure-limited frame processing tuned for on-device use

Hand tracking and gesture control

• 21-point hand landmark detection using MediaPipe Tasks
• Debounced, hold-based gestures to reduce accidental triggers
  • Pinch-hold — start / stop a session (or capture a calibration pose)
  • Open-palm-hold — switch exercise mode while paused
• On-screen controls (Start/Pause, End, mode selector) fully mirror the gestures, so the app stays usable if hand tracking is unavailable

Exercise modes

Bicep curls, squats, and push-ups — each with its own rep thresholds and independent calibration. Push-up mode shows an in-app hint to place the phone to your side (front-camera pose tracking of a plank is unreliable).

Form-gated rep counting

• Tracks rep phases and counts completed reps in real time
• Posture gating is phase-independent — it checks stability/visibility that holds across a rep (body in frame, level, torso not swinging), so counting is never paused mid-rep by the movement itself
• Low-quality movement doesn't get credited as a valid rep

Rep-quality feedback

• Direction-agnostic depth model (correct for curls, squats, and push-ups)
• Flags shallow reps and overly fast reps, with a per-rep score
• Maintains a recent rep timeline and a running average session score

Calibration

• Per-exercise calibration adapts thresholds to the user
• Stored locally with DataStore; reset-to-defaults supported

Session summary (graphs)

A scrollable dashboard rendered with hand-drawn Compose-Canvas charts:

• Average-score ring + total reps hero
• Score-per-rep line/area chart
• Verdict donut (good / shallow / too-fast) with legend
• Depth % and tempo small-multiple sparklines
• Full rep-quality timeline

Onboarding & developer tools

• First-run screen with a gesture how-to and on-device privacy note
• Built-in debug panel for inspecting thresholds, angles, posture, and calibration

Architecture

SmartForm uses a state-driven Compose UI backed by a ViewModel:

• SessionViewModel owns all in-session state (rep counting, posture, quality/session tallies, calibration) and exposes an immutable SessionUiState. It survives configuration changes and persists the selected mode via SavedStateHandle. Camera frames are fed in by plain method calls; hand frames land in a thread-safe StateFlow (they arrive on a background analyzer thread).
• SessionTracker is a pure, Android-free accumulator for per-session rep quality — unit-tested on the host JVM.
• RepCounter is a hysteresis + confirm-frame + EMA state machine with an injectable clock, so its transitions are deterministically testable.
• PoseMath holds the shared joint-angle geometry used by rep counting and posture evaluation.
• Frame pipeline: CameraX drives two ImageAnalysis use-cases on separate single-thread executors — PoseProcessor (ML Kit, YUV) and HandProcessor (MediaPipe, RGBA) — with STRATEGY_KEEP_ONLY_LATEST backpressure. Model initialization fails gracefully (a missing/corrupt hand model disables gestures instead of crashing).
• Charts are drawn by hand with Compose Canvas (ui/charts/) — no external charting dependency, fully theme-aware.

Tech stack

• Language: Kotlin
• UI: Jetpack Compose (Material 3), committed dark "neon" theme
• Architecture: MVVM (ViewModel + immutable UI state) with pure, testable domain logic
• Camera: CameraX
• Pose detection: ML Kit Pose Detection
• Hand tracking: MediaPipe Tasks Vision
• Local storage: DataStore Preferences
• Testing: JUnit (host-JVM unit tests)

Project structure

app/src/main/java/com/app/smartform/
├── calibration/
│   ├── CalibrationModels.kt
│   └── CalibrationStore.kt
├── camera/
│   └── CameraPreview.kt
├── gesture/
│   └── GestureDetector.kt
├── hand/
│   ├── HandModels.kt
│   ├── HandOverlay.kt
│   ├── HandProcessor.kt
│   └── YuvToRgbConverter.kt
├── pose/
│   ├── PoseFrame.kt
│   ├── PoseMath.kt
│   ├── PoseProcessor.kt
│   ├── PostureEvaluator.kt
│   └── SkeletonOverlay.kt
├── reps/
│   ├── ExerciseMode.kt
│   ├── RepCounter.kt
│   ├── RepQuality.kt
│   └── RepThresholds.kt
├── session/
│   ├── SessionStats.kt
│   ├── SessionTracker.kt
│   └── SessionViewModel.kt
├── ui/
│   ├── SessionSummaryScreen.kt
│   ├── charts/Charts.kt
│   └── theme/ (Color.kt, Theme.kt, Type.kt)
└── MainActivity.kt

Testing

Domain logic is covered by host-JVM unit tests:

./gradlew :app:testDebugUnitTest

• RepCounterTest — down/up transitions, hysteresis, form-gating, reset, arm selection
• RepQualityEvaluatorTest — depth %, verdicts, tempo thresholds, curl (inverted-threshold) depth
• SessionTrackerTest — verdict tallies, average score, timeline cap, snapshot
• PoseMathTest — joint-angle geometry
• CalibrationProfileTest — default thresholds

Requirements

• Android Studio with a current Android SDK (platforms 35+)
• Java 17+
• An Android device with a working camera (an emulator is not reliable for pose/hand validation)

Build and run

./gradlew :app:installDebug

Clean reinstall:

./gradlew :app:uninstallDebug
./gradlew :app:installDebug

Building a signed release

Release signing reads credentials from a git-ignored keystore.properties at the project root:

storeFile=/path/to/your.keystore
storePassword=********
keyAlias=********
keyPassword=********

Then:

./gradlew :app:assembleRelease

Notes:

• The release build type enables R8 minification and resource shrinking, with keep rules for ML Kit / MediaPipe (app/proguard-rules.pro).
• If keystore.properties is absent, release falls back to the debug key so local test builds still work.
• The output is a universal APK bundling native libraries for all ABIs, so it is large (~130 MB). This is fine for Uptodown; ABI splits can shrink it if needed.

Android configuration

• Min SDK: 26
• Target SDK: 35
• Compile SDK: 35
• Version: 1.0

Required permission:

<uses-permission android:name="android.permission.CAMERA" />

The MediaPipe hand-landmark asset lives at app/src/main/assets/hand_landmarker.task.

Limitations

• Gesture accuracy is best when the user is clearly visible and centered
• Very close distances reduce hand-landmark stability
• Low light reduces both pose and hand detection quality
• Push-ups are unreliable from a front camera; place the phone to your side
• The release APK is not yet 16 KB page-size compliant (a Google Play requirement for updates targeting Android 15+); it is unaffected on current devices and for Uptodown distribution

Troubleshooting

• Camera preview doesn't start — confirm camera permission, test on a physical device
• Gestures not recognized — keep the hand in frame, improve lighting, hold the gesture steadily; on-screen buttons always work as a fallback
• Reps not counting — check the form banner, make sure the selected mode matches the movement, or run calibration
• Calibration feels off — reset to defaults and recapture clean top/bottom poses

Roadmap

• Richer coaching cues during active reps
• Session history and trends across workouts
• Side-camera support for push-ups
• 16 KB page-size compliance for Play distribution
• Additional exercise modes and exportable summaries

Contributing

SmartForm is still evolving. Contributions that improve detection quality, exercise logic, UI clarity, performance, and documentation are welcome.