live_pose_detector

Real-time pose detection with a skeleton overlay, plus a built-in push-up form checker and rep counter — one widget, no permission boilerplate, no platform-channel code.

CameraPoseView owns the camera, permission, ML Kit detection, and skeleton rendering. PushUpFormDetector + PushUpFeedbackBanner add posture feedback and rep counting on top of the same landmark stream.

Requires Flutter >= 3.0, Dart >= 3.0, Android minSdk 21+, iOS 13+.

Platform	Supported
Android, iOS	✅
macOS, Windows, Linux, Web	❌ — google_mlkit_pose_detection (the ML backend) only ships Android + iOS

Install

dependencies:
  live_pose_detector: <latest_version>

Quick start

import 'package:live_pose_detector/live_pose_detector.dart';

Scaffold(
  body: CameraPoseView(),
)

That's the whole integration. Add your own logic via onPosesDetected:

CameraPoseView(
  initialLensDirection: CameraLensDirection.back,
  config: const PoseOverlayConfig(dotColor: Colors.greenAccent),
  onPosesDetected: (poses) {
    // your logic — same landmark stream, no extra detection work
  },
  onError: (error) {
    // optional — defaults to a built-in error view with Retry
  },
)

PoseOverlayConfig options

Field	Default	Description
dotColor / lineColor	green / yellow	Skeleton colors
dotRadius / lineWidth	5 / 3	Skeleton sizing
confidenceThreshold	0.5	Hides landmarks below this likelihood
connections	kPoseConnections	Skeleton topology — override to customize
resolutionPreset	medium	Camera quality vs. speed
detectionModel	accurate	ML Kit model quality vs. speed
showCameraSwitchButton	true	Built-in front/back FAB
requestCameraPermission	true	Set false if you handle permission yourself

Push-up form detection

final _detector = PushUpFormDetector();
final _feedback = ValueNotifier<PushUpFeedback?>(null);

CameraPoseView(
  onPosesDetected: (poses) => _feedback.value = _detector.evaluate(poses),
),
ValueListenableBuilder<PushUpFeedback?>(
  valueListenable: _feedback,
  builder: (context, feedback, _) => PushUpFeedbackBanner(feedback: feedback),
),

(Use a ValueNotifier, not setState — this fires every frame.)

evaluate() returns one PushUpFeedback per frame: a message, a color (success / warning / neutral), and the running rep count. Only one message is ever shown, picked safety-first: spine > neck > elbows > depth safety > tempo > insufficient depth > leg alignment > rep summary. Full list of states/messages/colors: PushUpState in lib/src/pushup/pushup_types.dart.

A rep only counts on a full correct cycle — top → controlled descent → correct depth → controlled ascent → top — with no form issue anywhere in between. Anything else finishes the cycle but reports incompletePushUp instead of counting.

Why it's not just raw angles vs. fixed thresholds: camera angle, body shape, and landmark noise all shift what "straight" looks like. So it:

• Calibrates each user's own neutral baseline from a genuinely stable top-position hold, not an assumed universal zero
• Widens tolerance near the bottom of a rep, since a 2D camera reads more deviation than actually exists as the torso foreshortens
• Debounces posture warnings (~200ms) so one noisy frame doesn't flash a false warning
• Gates on body orientation — if the body rotates away from the calibrated push-up angle (e.g. you stood up), rep tracking stops instead of counting nonsense

Every threshold is overridable via PushUpThresholds, and PushUpFeedback.debug exposes the raw numbers behind each judgment for tuning:

PushUpFormDetector(
  thresholds: const PushUpThresholds(minWarningPersistence: Duration(milliseconds: 300)),
)

Building your own features

onPosesDetected gives you every frame's List<Pose> (ML Kit's own types). Two helpers cover most rep-counting/gesture logic:

• angleBetweenLandmarks(a, center, b) — joint angle at center, e.g. elbow bend
• signedLineDeviation / signedLandmarkLineDeviation — signed distance of a point from a line, for straight-line checks (back straightness) where a plain angle can't tell which side something drifted to

See example/lib/pose_demo_screen.dart for both features wired to a real screen.

⚙️ Setup (required)

iOS — ios/Runner/Info.plist:

<key>NSCameraUsageDescription</key>
<string>Camera access is required to detect body pose.</string>

Deployment target 13.0+ in ios/Podfile.

Android — android/app/src/main/AndroidManifest.xml:

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

minSdkVersion 21+ in android/app/build.gradle.

Troubleshooting

Skeleton looks mirrored or offset — file an issue with device/orientation/lens details; this is unit-tested and shouldn't happen.

Feedback stuck on "Get into the push-up position" — no person detected, low landmark confidence, or the body-orientation gate tripped (you're not in a horizontal plank pose, or the camera moved).

Correct rep shows a warning (or vice versa) — camera setup, not a logic bug: go side-on, roughly hip height, whole body in frame, hold a straight top position for ~1s before starting. Still off? Check PushUpFeedback.debug; a consistent one-direction bias means try PushUpThresholds.invertBodyLineSign: true.

Warnings feel delayed — intentional (minWarningPersistence, default 200ms), filters landmark jitter. Lower it for faster but noisier feedback.

Known limitations

• No macOS/Windows/Linux/Web (ML Kit backend doesn't support them)
• No CI coverage for a real camera device — manual testing only
• Push-up detection expects one continuous, side-on camera view per set — not multi-angle footage
• Calibration assumes you start each set in a correct position; call PushUpFormDetector.reset() between sets

Example

Run example/ — a Start screen (camera picker) into a full-screen pose view with push-up feedback and rep counting wired up end to end.

Bugs & Credits

Report bugs and ask questions on GitHub Issues. Maintained by Dashstack Infotech, Surat.