What Is Pose Estimation for a Mobile Fitness App

Pose estimation locates specific points on a body in an image or a video. Those points are usually joints. A model returns coordinates — most often 2D [x, y] plus a confidence or visibility score — and a skeleton is just those points joined up. The model does not count a squat. Reps, joint angles and a form score are a product layer on top.

For a mobile fitness or wellness app in 2026–2027, with one person and a phone camera, start from MoveNet. Open MediaPipe (BlazePose) when you need 33 landmarks or 3D. Open YOLO pose when the job is many people or a model you export and run yourself. PoseTracker is the MoveNet product layer: on-device keypoints, then reps, angles and a form score, plus a public Assistant that turns a sentence into an integration.

Who already owns the query

Semrush, US, 5 October 2026. These are the pages that earn the visit. SDK marketing sites do not.

Query	US volume	Page that ranks	What that page is
pose estimation	880	Ultralytics pose task at #1	Model docs: definition, YOLO26-pose sizes, mAP, CPU and GPU speed
pose estimation models	210	Roboflow, best pose estimation models at #1	Roundup that recommends their own keypoint model
yolo pose	140	Same Ultralytics pose task, #1	Not a fitness SDK
mediapipe pose	480	Google Pose Landmarker at #1	Official MediaPipe docs
movenet	260	TensorFlow Hub MoveNet tutorial at #1	Official tutorial. The JS package is pose-detection
blazepose	140	Google research note at #1	The model inside MediaPipe Pose

The French Ultralytics URL (/fr/tasks/pose) does not earn traffic. The English task page does. PoseTracker does not publish a French mirror of that doc.

QuickPose, Sency, KinesteX and LightBuzz are real products. In this keyword set their organic traffic is brand, typos, or a thin landing — not the model head terms. Sency’s US traffic is almost entirely the query “sency”. KinesteX’s is mostly misspellings of the brand. LightBuzz ranks for body-tracking and Kinect-alternative queries, a different camera category.

Which stack, for a phone app

	Keypoints	Person	Camera	You still have to build
MoveNet (TensorFlow)	17 COCO	1	Phone, on-device	Reps, angles, UI
MediaPipe Pose / BlazePose	33, optional 3D	Configurable, often 1	Phone, on-device	The same, unless you wrap an SDK
YOLO pose (Ultralytics)	17 COCO on the pose models	Many, in one pass	Export (TFLite, CoreML, ONNX, TensorRT)	Dataset, export, reps
PoseTracker	17, MoveNet	1	Phone. React Native, iOS, Android, web	The screen. Keypoints are free. Reps need an API key
QuickPose	33, MediaPipe	Their SDK’s default	iOS-first, Android and React Native SDKs, on-device	Their feature set, not yours, where you call a built-in
KinesteX	Vendor skeleton	Their runtime	On-device SDK: React Native, Flutter, Swift, Kotlin, JS	Matching their white-label programs, or customizing them
Sency	Their Motion SDK	Their runtime	Mobile camera, demo-gated	A sales conversation before you integrate
LightBuzz	Their body model, 2D and 3D	Multi-person, including depth cameras	Phone, webcam, LiDAR, RealSense, Orbbec	A license. This is not a 17-point phone SDK

Phone FPS published for the browser path (TensorFlow.js, WebGL): MoveNet Lightning about 51 FPS on an iPhone 12 and 34 FPS on a Pixel 5. MediaPipe’s web build of BlazePose is about 11–12 FPS on that Pixel. Native MediaPipe on Android is a different, faster runtime. Full table: MediaPipe vs MoveNet. Eight models, including ML Kit, Apple Vision, OpenPose, MMPose and PoseNet: 2026 mobile guide.

Ultralytics publishes YOLO26n-pose at 57.2 mAP pose 50-95, 40.3 ms on CPU ONNX, 1.8 ms on an NVIDIA T4 with TensorRT. Those figures are from their pose task docs. They are not iPhone or Pixel FPS. Do not paste them into a phone budget.

What PoseTracker is, and is not

PoseTracker runs TensorFlow MoveNet on the device. Seventeen COCO keypoints. One person. A phone camera. An optional API key adds rep counting, joint angles and a form score. The Assistant on this site maps a catalog request to a snippet; custom movements continue after sign-in. It is a purpose-built integration assistant, not a general chatbot, and not a foundation model we trained from scratch.

It is not:

• a YOLO checkpoint, and not a replacement for the Ultralytics pose docs
• a MediaPipe or BlazePose fork (33 points, 3D)
• a depth-camera or multi-person tracker in the LightBuzz sense
• a white-label workout catalog in the KinesteX sense
• VRChat PoseTrackerVRC, marker-based 3D mocap, or an MMPose toolbox

Product comparison, including the public price shape: PoseTracker vs QuickPose vs KinesteX vs MediaPipe. We do not republish competitor price lists. QuickPose, Sency, KinesteX and LightBuzz change packaging; use their sites for a quote.

When to leave this page

• You want the model weights or the training loop. Use the official doc in the table above.
• You want reps in a React Native app next. Use pose estimation for fitness apps and the Assistant.
• You want range of motion from a camera, not a hardware goniometer. Read range of motion from the phone camera, then the flexibility tool.
• You want push-up reps inside an app. Read count push-ups from the phone camera.
• You want a clinical Modified Schober test with a tape. We don’t replace that. The back-flexibility tool is a side-view camera measure.