Case Study 03 · Edge Computer Vision & Machine Learning · 2023 – Present

Form Analyzer — On-Device Biomechanical ML Feedback

Rotation-invariant 3D Vector Triad Dot Product geometry and asynchronous inference gating on mobile devices.

Visit Live Platform View Source on GitHub View IEEE Paper Details →
Role & Responsibilities
Mobile & ML Developer / Researcher
Primary Stack
Flutter, Dart, Google ML Kit, Android NDK, 3D Vector Geometry
Key Engineering Metric
Zero-latency on-device processing maintaining 30–60 FPS with 0 server transmissions
Status & Verification
Published Research (IEEE ISEC 2026)
30–60 FPS On-Device Inference
0 ms Cloud Latency
33 Nodes 3D Spatial Tracking
IEEE 2026 Published Paper

Executive Overview

The Problem

Cloud-based fitness coaching apps suffer from 200–800ms latency, bandwidth costs, and privacy risks from uploading video feeds.

The Architecture

Edge Flutter client combining Google ML Kit pose detection with rotation-invariant 3D Vector Triad geometry and async frame gating.

Engineering Impact

Zero cloud latency, 60 FPS viewport smoothness, and verified rotation invariance across oblique camera angles (IEEE ISEC 2026).

[ Camera YUV420 Stream (60 FPS) ] ──► [ isBusy Inference Guard ]
                                                 │
                     ┌───────────────────────────┴───────────────────────────┐
                     ▼                                                       ▼
           [ Drop Frame (if Busy) ]                                [ Acquire Frame Lock ]
                                                                             │
                                                                             ▼
                                                                [ Google ML Kit 3D Poses ]
                                                                             │
                                                                             ▼
                                                                [ 3D Vector Triad Engine ]
                                                                - Direction Vectors u, v
                                                                - Dot Product Theta Solve
                                                                             │
                                                                             ▼
                                                                [ Biomechanical FSM Cues ]

Technical Architecture

1. 3D Vector Triad Dot Product Geometry

Evaluates interior joint angles from 3D spatial landmarks $(\vec{u} = P_1 - P_2, \vec{v} = P_3 - P_2)$, eliminating 2D planar foreshortening error:

\[\theta = \arccos\left( \frac{\vec{u} \cdot \vec{v}}{\|\vec{u}\| \|\vec{v}\|} \right)\]
3D Vector Triad Angle Engine Dart
```dart class VectorTriadCalculator { static double calculateJointAngle({ required PoseLandmark proximal, required PoseLandmark vertex, required PoseLandmark distal, }) { final double ux = proximal.x - vertex.x, uy = proximal.y - vertex.y, uz = proximal.z - vertex.z; final double vx = distal.x - vertex.x, vy = distal.y - vertex.y, vz = distal.z - vertex.z; final double dotProduct = (ux * vx) + (uy * vy) + (uz * vz); final double magU = math.sqrt(ux * ux + uy * uy + uz * uz); final double magV = math.sqrt(vx * vx + vy * vy + vz * vz); if (magU < 0.0001 || magV < 0.0001) return 180.0; final double cosTheta = (dotProduct / (magU * magV)).clamp(-1.0, 1.0); return math.acos(cosTheta) * (180.0 / math.pi); } } ```

2. Asynchronous Inference Gating (isBusy Guard)

Prevents backpressure by discarding intermediate camera frames during active neural net evaluation:

Asynchronous Frame Locking Service Dart
```dart class PoseDetectorService { final PoseDetector _detector = PoseDetector(options: PoseDetectorOptions(mode: PoseDetectionMode.stream)); bool _isBusy = false; void processCameraImage(CameraImage image, Function(List) onPoses) async { if (_isBusy) return; _isBusy = true; try { final inputImage = _convertCameraImage(image); final poses = await _detector.processImage(inputImage); if (poses.isNotEmpty) onPoses(poses); } finally { _isBusy = false; } } } ``` </div> </details> --- ## Key Takeaways & Verification - **Edge Privacy & Latency**: 100% on-device execution delivers instantaneous audio cues with 0 bytes transmitted off-device. - **Academic Publication**: Peer-reviewed and published at the 16th IEEE Integrated STEM Education Conference (ISEC 2026).