RepCounterSDK

Count exercise reps on-device, for free — no SDK key, no cloud.

RepCounterSDK turns Apple Vision body-pose landmarks into rep counts and form feedback. It's the open, MIT-licensed alternative to paid pose-fitness SDKs: everything runs on the Neural Engine, nothing leaves the phone, and there's no key to register.

import RepCounterSDK

let detector = RepDetector(exercise: .squat)

// Feed camera frames (e.g. from an AVCaptureSession), on a background queue:
switch detector.process(pixelBuffer) {
case .rep(let count):      print("reps:", count)
case .formIssue(let tip):  print(tip)            // e.g. "Bend both knees fully"
case .holding(let seconds): print("plank:", seconds)
case .idle: break
}

Why RepCounterSDK

	Paid pose SDKs	RepCounterSDK
On-device	✅	✅
SDK key / account required	❌ (key)	✅ none
Cost	💸	free (MIT)
Rep counter + form feedback	✅	✅
Anti-cheat (both sides must engage)	⚠️	✅
Extra model download / app size	⚠️	none (Apple Vision)

What makes the counting good

RepCounterSDK isn't a thin wrapper over landmarks — the rep logic is the point:

• Hysteresis counting — a rep only counts on a full rise past enter then fall past exit; the gap rejects frame jitter so noise can't double-count.
• Dual-joint anti-cheat — for symmetric moves (squat, push-up), each side has its own angle-gated counter, and a rep is only credited when both sides reached the bottom. Half-reps and one-sided cheating are rejected with a form tip.
• Angle-gated depth — reps register only when the relevant joint actually bent past a threshold (e.g. knees < 140°).
• Timed holds — planks accumulate seconds while the body stays straight.

Install

.package(url: "https://github.com/NazarKozak/RepCounterSDK.git", from: "0.1.0")

…and add "RepCounterSDK" to your target. Requires iOS 17+ / macOS 12+.

Exercises

Built in: .squat, .pushUp, .lunge, .bicepCurl, .shoulderPress, .plank.

Define your own (DSL)

Every exercise — built-in or custom — is an ExerciseSpec: a 0…1 progress signal plus optional anti-cheat gates. Define one in a few lines:

let deepSquat = ExerciseSpec.reps(
    name: "Deep squat",
    formTip: "Bend both knees fully",
    gates: [Gate.bent({ $0.kneeAngle($1) }, .left,  below: 120),
            Gate.bent({ $0.kneeAngle($1) }, .right, below: 120)],
    progress: { $0.depth(of: $0.kneeAngle, standing: 160, deep: 80) }
)

let plank = ExerciseSpec.hold(name: "Plank") { pose in
    pose.hipAngle().map { ($0 - 140) / 40 }   // straight body → ~1
}

let detector = RepDetector(spec: deepSquat)

Usage

From a camera pixel buffer (Apple Vision)

let detector = RepDetector(exercise: .pushUp)
// In your AVCaptureVideoDataOutput delegate (background queue):
let update = detector.process(sampleBuffer.imageBuffer!, orientation: .right)

process(_:) runs VNDetectHumanBodyPoseRequest synchronously — call it off the main thread.

From your own landmarks

Already have pose points (Vision, ARKit, a custom model)? Feed them directly:

let landmarks = PoseLandmarks(points: [
    .leftHip: hip, .leftKnee: knee, .leftAnkle: ankle, /* … */
])
let update = detector.process(landmarks)

RepCounterSDK only uses joint angles, so any consistent normalized coordinate space works.

API

• RepDetector(exercise:) — process(_:) -> RepUpdate, count, reset().
• RepUpdate — .rep(count:), .holding(seconds:), .formIssue(_), .idle.
• VisionPoseSource — standalone Apple Vision → PoseLandmarks.
• ThresholdCounter / ThresholdTimer — the reusable hysteresis primitives.
• PoseLandmarks — joint points + angle helpers (kneeAngle, elbowAngle, hipAngle).

Roadmap

• Squat, push-up, lunge (dual-joint anti-cheat), plank (timed hold)
• Bicep curl, shoulder press
• Apple Vision landmark source, free & on-device
• Hysteresis counter + threshold timer primitives
• Custom-exercise DSL (ExerciseSpec.reps/.hold)
• SwiftUI demo: live camera, skeleton overlay, rep counter
• Leg raise, jumping jack, sit-up
• Rep tempo / range-of-motion analytics

License

MIT — see LICENSE.