Problem
People who exercise at home often lack guidance on whether their posture and movement range are correct. Generic workout videos show what to do, but they do not evaluate the user's own performance.
AI-powered exercise form feedback from real workout videos.
PoseTrack helps home workout users compare their movement against a reference exercise video, receive understandable form feedback, and tracks improvement over time through scores, analytics, and gamification.
3Dpose analysis
Joint-Levelerror metrics
LLM + TTStext & spoken feedback
Product overview
What PoseTrack does
PoseTrack is a mobile fitness analysis application. Users upload a reference exercise video, record themselves performing the same movement, and receive both visual and textual feedback about how closely their form matches the reference.
Solution
PoseTrack turns video into structured motion data, aligns the user's movement with a reference sequence, computes joint-level differences, and presents the result as useful feedback, scores, and progress reports.
Demo area
Demo Video
This demo video demonstrates how our product is presented to the end-user. Click on the play button to see how PoseTrack works.
- Reference video import or upload
- User exercise recording
- Pose comparison result screen
- LLM-generated feedback shown & read aloud via text-to-speech
- Progress dashboard
Master features
Key features
Core algorithms & technologies
What powers PoseTrack
PoseTrack combines a Flutter mobile app, backend services, cloud storage, and a Python-based motion analysis pipeline for 3D exercise feedback.
3D human body model
SMPL-Based Pose Representation
Exercise videos are converted into structured 3D body motion data so the system can reason about body pose, joint rotations, and form deviations.
Analysis pipeline
Python Processing Modules
Frame extraction, pose output parsing, temporal alignment, scoring, and feedback payload generation are implemented in Python.
Model execution
PyTorch / GPU-Based Inference
Pose estimation models run in a Python + PyTorch environment, enabling heavier computation outside the mobile device.
Mobile application
Flutter Frontend
The app handles authentication, video input, camera recording, score screens, progress reports, and feedback visualization.
Video and result storage
AWS S3 Cloud Storage
Uploaded and recorded videos can be stored securely, while processed analysis outputs remain available for dashboards and history views.
Natural language feedback
Gemini LLM Feedback
Structured comparison results are converted into user-friendly explanations and actionable exercise advice.
Text-to-speech
ElevenLabs Voice Output
LLM-generated feedback is converted into natural-sounding speech using ElevenLabs, so users can hear corrections and coaching cues without taking their eyes off their form.
Frame ExtractionDecode videos into ordered analysis frames.
Temporal AlignmentMatch reference and user movement timing.
Rotation-Based ComparisonMeasure joint-level differences across matched frames.
Scoring & SeverityConvert metrics into accuracy scores and feedback priorities.
Technical pipeline
From videos to feedback
PoseTrack combines mobile UX, backend orchestration, GPU-based analysis, numerical comparison, and feedback generation.
1
Video input
Reference upload and user recording are collected from the app.
→
2
3D body recovery
SMPL-like joint rotations and body parameters are produced.
→
3
Alignment & comparison
Sequences are matched and compared with joint-level metrics.
→
4
Feedback
Scores, charts, visual highlights, and natural language advice are shown and read aloud via text-to-speech.
Comparison logic
The analysis module compares corresponding joints over aligned frames. Rotation-based differences, aggregate statistics such as mean/RMSE, and severity thresholds are used to detect which body parts need attention.
Feedback representation
Numerical outputs are structured as JSON so they can be consumed by the mobile UI, visualization layer, and optional LLM-based explanation module without tightly coupling the components.
System architecture
Component view
The end product is organized around a mobile client, backend services, storage, and an analysis pipeline that can run heavier computation outside the device.
Mobile App
Flutter UI · camera · video upload · feedback screens · progress dashboard
↓
Backend API
Authentication · video metadata · result retrieval · user history
Storage
User profiles · videos · processed outputs · performance records
↓
Preprocessing
Frame extraction and data normalization
Pose Analysis
3D pose estimation, temporal alignment, SO(3)-based comparison
Feedback Engine
Scoring, deviations, visual mappings, LLM text generation, text-to-speech audio output
FrontendFlutter, responsive mobile UI
BackendAPI endpoints, authentication, integration logic
AnalysisPython, PyTorch, pose estimation, comparison metrics
DataPersistent database, cloud/object storage
Project outcome
What was built
PoseTrack evolved from research and prototyping into an integrated mobile product with analysis and feedback components.
Research
3D pose estimation model study
Investigated SMPL-based models, output formats, GPU requirements, and alternatives when latency became a concern.
Mobile
Upload, recording, and user flow
Implemented core mobile flows for authentication, video input, camera usage, and result visualization.
Analysis
Pose comparison and scoring
Developed alignment, joint comparison, scoring, and structured output logic for visual and textual feedback.
Engagement
Progress and gamification
Added progress summaries, streak-oriented motivation, and dashboard elements to encourage consistent training.
Team
PoseTrack Members
Contact Us:
posetrackapp@gmail.comPoseTrack