Skip to content

PRD.md

Latest commit

 

History

History
130 lines (89 loc) · 7.86 KB

PRD.md

File metadata and controls

130 lines (89 loc) · 7.86 KB
Raw

PosturePro — Product Requirements Document

Status: Final Owner: Tanish Poddar One-liner: Real-time posture detection via computer vision — no backend, no login, just open and sit straight.

1. Problem

People who sit at desks for long stretches routinely develop forward head posture and hunched backs without realising it — there is no passive feedback while they work. Existing solutions either require wearable hardware or expensive software subscriptions. A lightweight, zero-friction tool that uses the webcam already built into every laptop and runs entirely in the browser can close that gap with no installation, no account, and no ongoing cost.

2. Goals (v1 / MVP)

  1. Detect posture in real time using only the device's built-in webcam — no additional hardware.
  2. Compute neck inclination and torso inclination angles geometrically from MediaPipe Pose landmarks.
  3. Classify each frame as good (neck < 40°, torso < 10°) or bad and display the verdict instantly on-screen.
  4. Track continuous bad-posture duration and fire a visible warning after 180 seconds.
  5. Ship a browser-based version with zero installation that works on any modern laptop.
  6. Ship a Python desktop version for users who need offline use or direct webcam control.
  7. Deployed with a working live demo URL (GitHub Pages).

3. Non-Goals (explicit scope cuts)

  • User accounts / sign-up — The tool is anonymous; no identity management is needed for v1.
  • Posture history / analytics — Session data is not persisted; the page resets on reload. Deferred to v2 when there is a clear user demand for trends.
  • OS-level or audio notifications — Alerts are visual only (on-screen banner). System notifications and sound cues deferred until browser notification APIs are evaluated for cross-platform support.
  • Mobile support — Portrait-orientation body tracking requires different landmark extraction; deferred to v2.
  • Per-user calibration — Fixed angle thresholds apply universally. Personalised baselines require a calibration flow that adds friction for v1.
  • Cloud processing or video upload — All inference runs on-device; no video ever leaves the browser.

4. Users

Primary: Students, developers, and desk workers who sit for long periods and want passive real-time feedback on their posture without installing software or creating an account.

Secondary: Recruiters and engineers reviewing this as a portfolio project — the live demo must work reliably on first visit with a single button click.

5. User Stories

  1. As a desk worker, I want to open a URL and start posture monitoring in under 10 seconds so that I can get feedback without any setup.
  2. As a user, I want to see my neck angle, torso angle, and shoulder alignment update live so that I understand exactly what the system is measuring.
  3. As a user, I want a clear colour-coded indicator (green/red) on the video overlay so that I can glance at my screen and know my posture status immediately.
  4. As a user, I want a prominent warning banner after I have had bad posture for 3 minutes so that I am prompted to correct my position before pain sets in.
  5. As a developer, I want a Python desktop app I can run offline so that I can use the tool without a browser or internet connection.
  6. As a recruiter, I want the live demo to detect my posture within 2 seconds of granting camera access so that the project's functionality is immediately obvious.

6. Functional Requirements

6.1 Camera Access

  • The web app must request camera access via getUserMedia with { width: 640, height: 480, facingMode: 'user' }.
  • If camera access is denied, a clear error message must be displayed.
  • The user must be able to stop the camera feed with a single button click.

6.2 Pose Detection

  • The system must use MediaPipe Pose (model complexity 1) to detect body landmarks on each frame.
  • The system must extract four specific landmarks: left shoulder, right shoulder, left ear, and left hip.
  • If no pose is detected in a frame, the system must display a "no pose" indicator rather than crashing.

6.3 Posture Analysis

  • The system must compute shoulder offset (Euclidean distance between left and right shoulder landmarks).
  • The system must compute neck inclination (angle between left shoulder → left ear vectors relative to vertical).
  • The system must compute torso inclination (angle between left hip → left shoulder vectors relative to vertical).
  • Posture must be classified as good if neck inclination < 40° AND torso inclination < 10°, otherwise bad.

6.4 Feedback & Alerts

  • The skeleton overlay (landmark dots and connecting lines) must be drawn on a <canvas> overlay in green for good posture and red for bad posture.
  • Live metric values (neck°, torso°, shoulder offset, posture timer) must be displayed in a stats panel below the video.
  • The posture timer must show how long the current good or bad streak has lasted.
  • A warning banner must appear when the user has been in bad posture for more than 180 consecutive seconds.

6.5 Python Desktop App

  • The Python app must capture from a webcam using OpenCV and process frames with the MediaPipe Python SDK.
  • The app must display the annotated video feed in a live OpenCV window.
  • The app must write the annotated session to output.mp4 for review.
  • The app must print a console warning after 180 seconds of bad posture.

7. Non-Functional Requirements

  • Latency: Pose inference must run at a framerate sufficient for smooth visual feedback (target ≥ 15 fps on a mid-range laptop).
  • Privacy: No video frames or biometric data may leave the device — all processing is local (WASM in browser, native in Python).
  • Cost: Zero server costs. All inference runs on the client; hosting is GitHub Pages (free).
  • Compatibility: The web app must work in Chrome and Edge (desktop); both ship MediaPipe-compatible WASM runtimes.
  • No dependencies on sign-up or API keys: The app must function with no credentials of any kind.

8. Success Metrics

Metric Target
Time from page load to first posture reading < 10 seconds on a typical laptop
Warning fires correctly After exactly 180 seconds of continuous bad posture
Live demo reliability Works on first visit without refresh in Chrome
Python app launch-to-detection < 5 seconds from running the script

9. Risks & Open Questions

  • MediaPipe CDN availability — The web app depends on jsDelivr to serve MediaPipe WASM files. If the CDN is unavailable, the app cannot load. Mitigated by jsDelivr's high uptime SLA; a v2 fix would bundle the WASM locally.
  • Camera index (Python) — The desktop app hardcodes cv2.VideoCapture(1), which fails if the webcam is at index 0. Mitigated by a clear error message; v2 should auto-detect or prompt.
  • Lighting conditions — MediaPipe Pose detection degrades significantly in poor lighting. Documented in setup notes; no algorithmic mitigation planned for v1.
  • Angle formula edge cases — The arccos formula produces NaN if the landmark coordinates produce a value outside [-1, 1]; wrapped in try/except (Python) and guarded by early return (JS).

10. v2 Candidates

  • Posture history chart — Session-level posture timeline stored in localStorage, visualised as a sparkline so users can track improvement over time.
  • OS / audio notifications — System-level desktop notifications and an optional sound alert so warnings reach users who minimise the browser tab.
  • Per-user calibration — A 10-second guided calibration flow that captures the user's natural upright posture as the baseline, replacing fixed 40°/10° thresholds.
  • Mobile support — Portrait-mode landmark extraction and a responsive layout designed for phone cameras.
  • Bundled WASM — Ship MediaPipe assets with the repo so the app works without a CDN dependency.